AIDb — The AI Database: every AI model, indexed

AIDb is the IMDb for AI models — a verified, source-backed directory of 1176 AI models across 329 vendors. Browse large language models, reasoning models, image, video and music generators, speech, vision and robotics models from OpenAI, Google DeepMind, Anthropic, Meta AI, Mistral AI, Alibaba, DeepSeek, xAI and more. Each model has a full dossier covering capabilities, vendor, jurisdiction, pricing, open-source status and verified release data with primary sources. Select any model below to open its dossier.

• Yi-1.5 01.AI · LLM · 2024 — 01.AI's Yi-1.5, an upgraded Yi series open-sourced on May 13, 2024 in 6B, 9B and 34B sizes (base and chat, with 4K, 16K and 32K context versions), continually pretrained on Yi with 500B more high-quality tokens (3.6T in total) and fine-tuned on 3M samples; 01.AI reports stronger coding, math, reasoning and instruction following than Yi while keeping its language understanding, commonsense reasoning and reading comprehension. Apache 2.0.
• Yi-34B 01.AI · LLM · 2023 — 01.AI's Yi-34B, the larger of two bilingual (English and Chinese) base models in Yi's first open release on November 2, 2023, trained from scratch on 3T tokens using the Llama architecture but not Llama's weights, with a 4K sequence length extendable to 32K at inference; 01.AI says it ranked first among open-source models in both English and Chinese on benchmarks including the Open LLM Leaderboard and C-Eval as of November 2023, and Yi-34B-Chat placed second to GPT-4 Turbo on AlpacaEval.
• Yi-6B 01.AI · LLM · 2023 — 01.AI's Yi-6B, the smaller of two bilingual (English and Chinese) base models in Yi's first open release on November 2, 2023, trained from scratch on 3T tokens using the Llama architecture but not Llama's weights, with a 4K sequence length extendable to 32K at inference; 01.AI suggests the 6B models for personal and academic use, and a 200K-context Yi-6B-200K and Yi-9B, trained further from Yi-6B on 0.8T tokens, followed.
• Yi-Coder 01.AI · Code LLM · 2024 — 01.AI's Yi-Coder, a series of open code language models released September 5, 2024 in 1.5B and 9B sizes (base and chat), with a 128K-token context and support for 52 programming languages; 01.AI reports Yi-Coder-9B-Chat reached a 23% pass rate on LiveCodeBench, the only model under 10B parameters above 20%, ahead of DeepSeek-Coder-33B-Instruct at 22.3%. Apache 2.0.
• Yi-Large 01.AI · LLM · 2024 — 01.AI's Yi-Large, offered through the 01.AI API Platform with a 32K context window at $3 per million input and output tokens; 01.AI calls it a state-of-the-art model for chat and text generation, aimed at complex inference, prediction and natural-language understanding, and the platform also lists Yi-Large-Turbo (4K context) and Yi-Large-FC, a variant for tool use.
• Yi-Lightning 01.AI · LLM · 2024 — 01.AI's Yi-Lightning, a flagship mixture-of-experts LLM released in October 2024 that 01.AI describes as a 100B-parameter model, with fine-grained expert segmentation and routing, optimized KV caching and a context extended to 64K tokens; trained with pre-training, supervised fine-tuning and RLHF, it ranked sixth overall on Chatbot Arena, placing second to fourth in Chinese, Math, Coding and Hard Prompts, and is served through 01.AI's developer platform.
• Fuyu-8B Adept · Multimodal LLM · 2023 — Adept's Fuyu-8B, an open small version of the multimodal model behind Adept's product: a plain decoder-only transformer with no separate image encoder that feeds image patches straight into its first layer, so it takes images of any resolution; built for digital agents, it answers questions about charts, diagrams and user interfaces, locates items on screen images and responds to large images in under 100 milliseconds. A base model meant for fine-tuning; CC BY-NC 4.0.
• Adobe Firefly 1 Adobe · Image Generation · 2023 — Adobe's first Firefly model, introduced in beta on March 21, 2023 as the start of its Firefly generative AI family, focused on images and text effects; Adobe trained it on hundreds of millions of licensed Adobe Stock images plus openly licensed and expired-copyright public domain content so its output would suit commercial use, added a style engine, and backed Content Credentials and a 'Do Not Train' tag. Users made over 3 billion images with it before Firefly Image 2 arrived in October 2023.
• Adobe Firefly 2 Adobe · Image Generation · 2023 — Adobe's Firefly Image 2 Model, unveiled at Adobe MAX on October 10, 2023 with the Firefly Vector and Design models: it raised image quality, improved rendering of skin, hair, eyes, hands and bodies, colors and dynamic range, and added Generative Match for user-chosen reference styles, camera-like Photo Settings and Prompt Guidance. It was trained on licensed content such as Adobe Stock and public domain content whose copyright has expired, and it accepts prompts in over 100 languages.
• Adobe Firefly 3 Adobe · Image Generation · 2024 — Adobe's Firefly Image 3 Foundation Model, debuted in beta on April 23, 2024 in Photoshop (beta) and the Firefly web app, with better photo quality (lighting, people, crowds), prompt understanding, detail, variety and text rendering, a new style engine for auto-stylization, and Structure Reference and Style Reference controls. Like earlier Firefly models it was trained on licensed content such as Adobe Stock and designed for commercial use, and Adobe offers IP indemnification for Firefly output.
• Adobe Podcast Adobe · Audio Processing · 2023 — Adobe Podcast, Adobe's browser-based tool for 'AI audio recording and editing, all on the web': Enhance Speech removes noise and echo from voice recordings, including in video files, and Studio (beta) records solo or with remote guests on separate tracks, edits audio like a document through its transcript, and adds captions and audiograms, alongside transcription and music-removal tools. A free plan is available, and Premium features are included with the Adobe Firefly Pro plan.
• mGPT AI Forever/Sber · LLM · 2022 — SberDevices' mGPT (published by AI Forever), a multilingual GPT-3-style language model with 1.3B parameters trained on 61 languages from 25 language families using Wikipedia and the C4 corpus (about 600 GB of text, 440 billion tokens) with sparse attention on Megatron and DeepSpeed; it performs on par with Meta's XGLM while covering more languages, including low-resource languages of CIS countries and smaller peoples of Russia, and a 13B version was also trained. Apache 2.0.
• SEA-LION AI Singapore · LLM · 2023 — AI Singapore's SEA-LION (Southeast Asian Languages in One Network), a family of open models built for Southeast Asia's languages and cultures: v1 (3B and 7B) was pre-trained from scratch, later versions continue pre-training or fine-tune Llama, Gemma, Qwen and NVIDIA Nemotron bases, v4 added image input with up to 256K context, and the September 2026 v4.8 brought its first mixture-of-experts models and coverage of 13 regional languages, alongside SEA-LION embedding and SEA-Guard safety models.
• Jamba AI21 Labs · LLM · 2024 — AI21's original Jamba base model: a hybrid Mamba (SSM) and Transformer mixture-of-experts LLM with 12B active and 52B total parameters and a 256K context (up to 140K tokens on one 80GB GPU), knowledge cutoff March 5, 2024. AI21 has since released the instruction-tuned Jamba 1.5 Mini and Large.
• Jamba 1.5 AI21 Labs · LLM · 2024 — AI21 Labs' Jamba 1.5, instruction-tuned open models built on the Jamba hybrid Transformer-Mamba mixture-of-experts architecture for high throughput and low memory use: Jamba 1.5 Large (94B active, about 399B total parameters) and Jamba 1.5 Mini (12B active), both with an effective 256K-token context, the longest among open-weight models at release; ExpertsInt8 quantization fits Large on eight 80GB GPUs at full context. Jamba Open Model License.
• Jurassic-1-Jumbo AI21 Labs · LLM · 2021 — AI21's 178B-parameter model (Aug 2021), at release the largest language model available to developers for general use; it uses a 250K-token vocabulary that includes multi-word tokens.
• Jurassic-2 AI21 Labs · LLM · 2023 — AI21's 2023 model family in three sizes (Large, Grande and Jumbo), plus instruction-tuned Jumbo and Grande versions, offered through AI21 Studio.
• Wordtune AI21 Labs · AI Writing · 2020 — AI rewriting and summarization.
• AIVA AIVA · Music Generation · 2016 — AI composer. Classical and film music.
• Aleph Alpha Luminous Aleph Alpha · LLM · 2021 — European sovereign AI model.
• Luminous Supreme Aleph Alpha · LLM · 2022 — 70B European model. Multilingual.
• Babel-83B Alibaba · LLM · 2025 — Multilingual LLM covering the 25 most-spoken languages, which together serve over 90% of the world's population; the 83B-parameter flagship of the Babel family.
• GTE-Large Alibaba · Embeddings · 2023 — Alibaba DAMO Academy's GTE-large (General Text Embeddings), a BERT-based English text embedding model producing 1,024-dimensional vectors for inputs of up to 512 tokens, trained with multi-stage contrastive learning on a large corpus of relevance text pairs for retrieval, semantic textual similarity and reranking; it averaged 63.13 across 56 MTEB datasets in DAMO's comparison, ahead of e5-large-v2 and OpenAI's text-embedding-ada-002. MIT license.
• ModelScope T2V Alibaba · Video Generation · 2023 — Alibaba's ModelScope text-to-video model (ModelScopeT2V), which evolves Stable Diffusion's text-to-image design with spatio-temporal blocks for consistent frames and smooth motion; its VQGAN, text encoder and denoising UNet total 1.7 billion parameters, 0.5 billion of them for temporal modeling, and it handles varying frame counts. Code and a demo are on ModelScope.
• ParScale Alibaba · LLM · 2025 — Research checkpoints for parallel scaling (ParScale): running several learned parallel streams through one model and combining them, a way to scale beyond adding parameters or inference time. Academic-use checkpoints, Apache 2.0.
• PolyLM Alibaba · LLM · 2023 — PolyLM, a polyglot decoder-only LLM in 1.7B and 13B sizes, trained on 640B tokens with bilingual data and a curriculum that raises the non-English share from 30% to 60%; it covers 18 languages and, per its authors, beats LLaMA and BLOOM on multilingual tasks while staying comparable in English. An instruction-tuned version used 132.7K self-generated multilingual instructions.
• Qwen Alibaba · LLM · 2023 — Alibaba Cloud's Qwen (Tongyi Qianwen), the first generation of its open LLM series: base models Qwen-1.8B, Qwen-7B, Qwen-14B and Qwen-72B, released between August and November 2023 and pretrained on 2.2 to 3 trillion tokens of multilingual data focused on Chinese and English, plus Qwen-Chat models aligned with human preferences that can chat, write, summarize, translate, code, solve math problems, use tools, act as agents and work as code interpreters.
• Qwen 1.5 Alibaba · LLM · 2024 — Alibaba's Qwen1.5, the beta version of Qwen2, first released in February 2024 as open base and chat models from 0.5B to 110B parameters plus a 14B mixture-of-experts model with 2.7B activated, all with a 32K context; the Qwen team reports better alignment with human preferences and stronger multilingual ability than the first Qwen, with Qwen1.5-72B beating Llama 2 70B on every benchmark it listed, and native Hugging Face Transformers support.
• Qwen-72B Alibaba · LLM · 2023 — Alibaba Cloud's Qwen-72B, the 72-billion-parameter base model of the first Qwen (Tongyi Qianwen) series, released with the Qwen-72B-Chat assistant; pretrained on over 3 trillion tokens of Chinese, English, multilingual, code and math data, it has a vocabulary of over 150K tokens and a 32K context, and Alibaba reports it significantly surpassed existing open-source models on Chinese and English benchmarks. Commercial use requires an application under the Tongyi Qianwen license.
• Qwen-Audio Alibaba · Audio LLM · 2023 — Alibaba Cloud's Qwen-Audio, the audio-language model of the Qwen series, which takes diverse audio (human speech, natural sounds, music and songs) plus text as input and outputs text; a multi-task training framework covering more than 30 tasks gives it strong results without task-specific fine-tuning, including state-of-the-art scores on Aishell1, CochlScene, ClothoAQA and VocalSound, and the instruction-tuned Qwen-Audio-Chat adds multi-turn audio and text dialogue.
• Qwen-Image Alibaba · Image Generation · 2025 — Qwen's first image-generation foundation model: a 20B MMDiT (Apache 2.0) notable for native text rendering, including Chinese, and consistent image editing; uses Qwen2.5-VL for semantic encoding plus a VAE path.
• Qwen-VL Alibaba · Multimodal LLM · 2023 — Alibaba Cloud's Qwen-VL, the vision-language version of Qwen, built from Qwen-7B and an OpenCLIP ViT-bigG encoder; it takes images, text and bounding boxes as input and outputs text and boxes, handles multilingual and multi-image conversation, reads Chinese and English text in images and uses 448x448 input for fine-grained recognition. Released openly with Qwen-VL-Chat in August 2023, it beat open models of similar scale on captioning, VQA, DocVQA and grounding.
• Qwen-VL-Max Alibaba · Multimodal LLM · 2024 — Alibaba's Qwen-VL-Max, the more capable of two upgraded Qwen-VL vision-language models presented in January 2024 with Qwen-VL-Plus, with stronger visual reasoning and instruction following, better reading of image details and text, and support for images above one million pixels; Alibaba reported results on par with Gemini Ultra and GPT-4V on several text-image tasks and ahead of both on Chinese question answering and text comprehension. Offered through DashScope APIs and the Qwen website.
• Qwen-VL-Plus Alibaba · Multimodal LLM · 2024 — Alibaba's Qwen-VL-Plus, which the Qwen team calls its enhanced large vision-language model, one of two upgraded Qwen-VL models presented in January 2024 with the more capable Qwen-VL-Max; it is significantly better at recognizing details and text in images and accepts images of over one million pixels at any aspect ratio, and Alibaba reported the two perform on par with Gemini Ultra and GPT-4V on several text-image tasks. Offered through DashScope APIs and the Qwen website.
• Qwen2-VL Alibaba · Multimodal LLM · 2024 — Alibaba's Qwen2-VL, the second-generation Qwen vision-language series (2B, 7B and 72B), which handles images at any resolution through Naive Dynamic Resolution and fuses text, image and video positions with Multimodal Rotary Position Embedding; it understands videos over 20 minutes long, reads multilingual text in images, can act as an agent operating phones or robots, and posted state-of-the-art results on benchmarks such as MathVista and DocVQA. The 2B and 7B models are Apache 2.0.
• Qwen2.5-72B Alibaba · LLM · 2024 — Alibaba's Qwen2.5-72B, the largest open model in the Qwen2.5 series: a 72.7B-parameter dense transformer (80 layers, grouped-query attention) with a 131,072-token context and up to 8,192 tokens of output, released as base and instruction-tuned models; compared with Qwen2 it has much more knowledge, stronger coding and math, better instruction following, long-text and JSON generation, and support for over 29 languages. Released under the Qwen license.
• Qwen2.5-Coder Alibaba · Code LLM · 2024 — Alibaba's Qwen2.5-Coder, the code-specific Qwen2.5 series that succeeded CodeQwen1.5, in six sizes from 0.5B to 32B parameters, each as base and instruct models, continually pretrained from Qwen2.5 on over 5.5 trillion tokens, with a 32K context up to 3B and 128K from 7B; Alibaba reports state-of-the-art open-model results on more than 10 code benchmarks, with the 32B instruct model matching GPT-4o at coding at its November 2024 release. All sizes except 3B are Apache 2.0.
• Qwen2.5-Coder-32B Alibaba · Code LLM · 2024 — Alibaba's Qwen2.5-Coder-32B, the largest model in the Qwen2.5-Coder code series: a 32.5B-parameter model (64 layers) built on Qwen2.5 and trained on 5.5 trillion tokens of source code, text-code grounding and synthetic data, with a 131,072-token context; Alibaba says the instruct version became the state-of-the-art open-source code LLM at release, with coding abilities matching GPT-4o, while keeping strong math and general skills. Apache 2.0.
• Qwen2.5-Max Alibaba · LLM · 2025 — Large-scale mixture-of-experts model pretrained on over 20 trillion tokens and post-trained with supervised fine-tuning and RLHF (Jan 2025); offered through Alibaba Cloud's API and Qwen Chat.
• Qwen3-0.6B Alibaba · LLM · 2025 — Dense 0.6B-parameter model (0.44B non-embedding) in Alibaba's Qwen3 family, with 28 layers and a 32K context. Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-1.7B Alibaba · LLM · 2025 — Dense 1.7B-parameter model in Alibaba's Qwen3 family, with 28 layers and a 32K context. Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-14B Alibaba · LLM · 2025 — Dense 14.8B-parameter model in Alibaba's Qwen3 family, with 40 layers and a 32K native context (131K with YaRN). Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-235B-A22B Alibaba · LLM · 2025 — Flagship mixture-of-experts model of Alibaba's Qwen3 family: 235B total and 22B active parameters (128 experts, 8 active), 94 layers and a 32K native context (131K with YaRN), pretrained on about 36 trillion tokens covering 119 languages and dialects. It switches between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, and supports tool calling.
• Qwen3-30B-A3B Alibaba · LLM · 2025 — Mixture-of-experts model in Alibaba's Qwen3 family with 30.5B total and 3.3B active parameters (128 experts, 8 active), 48 layers and a 32K native context (131K with YaRN). Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-32B Alibaba · LLM · 2025 — Dense 32.8B-parameter model in Alibaba's Qwen3 family, with 64 layers and a 32K native context (131K with YaRN). Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-4B Alibaba · LLM · 2025 — Dense 4.0B-parameter model in Alibaba's Qwen3 family, with 36 layers and a 32K native context (131K with YaRN). Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-8B Alibaba · LLM · 2025 — Dense 8.2B-parameter model in Alibaba's Qwen3 family, with 36 layers and a 32K native context (131K with YaRN). Qwen3 models switch between a thinking mode for math, coding and logical reasoning and a non-thinking mode for general chat within one model, support tool calling and cover 100+ languages and dialects.
• Qwen3-Coder Alibaba · Code LLM · 2025 — Qwen3-Coder, the Qwen team's open-weight (Apache 2.0) agentic coding family, led by Qwen3-Coder-480B-A35B-Instruct: a 480B mixture-of-experts model with 35B active, a 256K-token native context (1M with YaRN), 7.5 trillion pretraining tokens (70% code) and code and agent RL across 20,000 parallel environments, which Qwen calls comparable to Claude Sonnet 4 on agentic coding. It launched with the Qwen Code CLI and a Model Studio API; 30B-A3B and Qwen3-Coder-Next versions followed.
• Qwen3-Max Alibaba · LLM · 2025 — Alibaba's largest LLM when unveiled at its Apsara Conference in Sep 2025, with over 1 trillion parameters; closed weights. It was offered through Qwen Chat and the Alibaba Cloud Model Studio API, whose price list no longer includes it; Alibaba's Qwen3.8-Max (2.4 trillion parameters) is now larger.
• Qwen3-Max-Preview Alibaba · LLM · 2025 — Preview of Qwen3-Max released Sep 5, 2025: Alibaba's first model with over 1 trillion parameters. Closed weights, available through Qwen Chat and the Alibaba Cloud API.
• Qwen3-Next-80B-A3B Alibaba · LLM · 2025 — First Qwen3-Next model: an 80B-parameter MoE with only 3B active, using hybrid attention (Gated DeltaNet plus gated attention) for efficient ultra-long context. Open weights (Apache 2.0).
• Qwen3-Omni Alibaba · Multimodal LLM · 2025 — Natively end-to-end omni-modal model: takes text, images, audio and video and streams answers as text or natural speech; supports 119 text languages. Open weights (Apache 2.0).
• Qwen3-VL Alibaba · Multimodal LLM · 2025 — Qwen's third-generation vision-language family. Flagship open-weights Qwen3-VL-235B-A22B (MoE; Instruct and Thinking editions) with 256K native context (extendable to 1M), long-video understanding, OCR in 32 languages, 2D/3D grounding and GUI-agent control.
• Qwen3.5 Alibaba · LLM · 2026 — First open-weight model of Alibaba's Qwen3.5 series: Qwen3.5-397B-A17B, a native vision-language model whose hybrid architecture combines linear attention (Gated Delta Networks) with a sparse mixture of experts, 397B total and 17B active parameters; supports 201 languages and dialects, up from 119.
• Qwen3.6 Plus Alibaba · LLM · 2026 — Alibaba's hosted Qwen3.6 flagship, available through Alibaba Cloud Model Studio's API, with a 1M-token context window by default and stronger agentic coding and multimodal reasoning than Qwen3.5-Plus; the smaller Qwen3.6 models are open weights.
• Qwen3.6-35B-A3B Alibaba · LLM · 2026 — Qwen3.6-35B-A3B, the first open-weight Qwen3.6 model (Apache 2.0, April 2026): a mixture-of-experts language model with a vision encoder, 35B total and 3B active parameters, hybrid Gated DeltaNet and gated attention layers, and a 262K-token native context extendable to about 1M. Built for agentic coding, it can keep reasoning from earlier turns; Qwen reports 73.4 on SWE-bench Verified, 51.5 on Terminal-Bench 2.0, 86.0 on GPQA and 81.7 on MMMU.
• Qwen3.7-Max Alibaba · Reasoning LLM · 2026 — Qwen3.7-Max, Alibaba's closed-weight Qwen3.7 flagship, unveiled at the Alibaba Cloud Summit on May 20, 2026 for agentic coding, office work and long-horizon tasks: Alibaba says it can run autonomously for up to 35 hours and over 1,000 tool calls without degrading and is tuned for agent frameworks such as Claude Code, OpenClaw and Qoder. Its API took text only at launch (a June 8 snapshot added vision), with thinking on by default, a 1M-token context and up to 131K output tokens.
• Qwen3.8-27B Alibaba · Multimodal LLM · 2026 — Qwen3.8-27B, Alibaba's open-weight (Apache 2.0) dense 27B vision-language model in the Qwen3.8 family, built on Qwen3.6-27B with better coding and office work across text and images: it understands images and hour-long videos, thinks by default with adjustable reasoning effort, and has a 262K-token native context extendable to 1M. Qwen reports 73.0 on Terminal Bench 2.1, 61.7 on SWE-bench Pro and 89.2 on GPQA Diamond; Qwen Cloud also hosts it at $0.50/$3 per million tokens.
• Qwen3.8-Flash-Next Alibaba · Multimodal LLM · 2026 — Qwen3.8-Flash-Next (August 26, 2026), Alibaba's open-weight (Qwen Community License) multimodal mixture-of-experts model with 125B parameters (6B active) plus 51B of n-gram embeddings, released as an experimental preview of the architecture that will underpin Qwen4: Qwen Sparse Attention, gated residual streams, n-gram embeddings and a Muon and AdamW training recipe. Its card reports 62.5 on SWE-bench Pro, 91.7 on GPQA Diamond and 91.9 on LiveCodeBench v6; the hosted version is Qwen3.8-Flash.
• Qwen3.8-Max Alibaba · Multimodal LLM · 2026 — Qwen3.8-Max, Alibaba's flagship native vision-language mixture-of-experts model (2.4T total, 95B active parameters), built on Qwen3.8-2.4T-A95B, which Qwen released as its first open Qwen-Max-class model. The API version adds vision input, a non-thinking mode, a 1M-token context and built-in tools, and Alibaba says it can code autonomously for 10+ days; Qwen's card reports 86.6 on Terminal Bench 2.1, 67.7 on SWE-bench Pro, 92.6 on GPQA Diamond and 93.0 on PaperBench.
• Qwen3.8-Max-0902 Alibaba · Multimodal LLM · 2026 — Qwen3.8-Max-0902 (alias qwen3.8-max-2026-09-02), the September 2, 2026 snapshot of Alibaba's 2.4-trillion-parameter Qwen3.8-Max: Alibaba says it handles more complex engineering-scale projects and long-horizon autonomous development, orchestrates tools more reliably, and sharpens chart reasoning, document parsing and multimodal perception, while keeping the 1M-token context, thinking mode and tool ecosystem. From September 5, 2026 the qwen3.8-max API name switched to it at unchanged prices.
• QwenLong-L1-32B Alibaba · LLM · 2025 — Long-context reasoning model trained with reinforcement learning for document-level reasoning, from Alibaba's Tongyi team. Apache 2.0.
• QwQ-32B Alibaba · Reasoning LLM · 2025 — Medium-sized reasoning model of the Qwen series (32B), which Qwen says is competitive with DeepSeek-R1 and o1-mini. Apache 2.0.
• QwQ-Max-Preview Alibaba · Reasoning LLM · 2025 — Preview of Qwen's reasoning model built on Qwen2.5-Max, aimed at deep reasoning, math, coding and agent tasks (Feb 2025).
• Tongyi Qianwen Alibaba · Chat Interface · 2023 — Alibaba's Qianwen AI assistant app, formerly Tongyi Qianwen, whose tongyi.com address now redirects to qianwen.com; Alibaba describes it as its official AI assistant and the first place to use its strongest Qwen models, with AI search, web-page summaries, AI slide (PPT) creation, AI image generation and notes from audio recordings.
• WAN Alibaba · Video Generation · 2025 — Wan2.1, Alibaba's open (Apache 2.0) suite of video foundation models, first released February 25, 2025 with 14B and 1.3B text-to-video and 14B image-to-video models, later joined by first-and-last-frame and VACE creation-and-editing models. Its team says it beats open-source and commercial models on several benchmarks, the 1.3B model needs only 8.19 GB of VRAM (a 5-second 480p clip in about 4 minutes on an RTX 4090), and it was the first video model to render both Chinese and English text.
• Wan 3.0 Alibaba · Video Generation · 2026 — Wan 3.0, Alibaba's all-in-one video generation model, launched by invitation on August 6, 2026 and generally available from August 24: one model handles text-to-video, image-to-video from first or first-and-last frames, reference-to-video, editing, replication and driving, takes text, image, audio and video references together, and makes videos up to 30 seconds long with sound. A faster Wan3.0-Video-Prime version is also offered; API prices start at $0.05 a second for 480p.
• Qwen2 Alibaba Cloud · LLM · 2024 — Alibaba's Qwen2, released June 7, 2024 as pretrained and instruction-tuned models in five sizes (0.5B, 1.5B, 7B, a 57B-A14B mixture-of-experts model and 72B), all using grouped-query attention, trained on data in 27 languages beyond English and Chinese, and with contexts up to 128K tokens for the 7B and 72B instruct models; the Qwen team reports state-of-the-art results on many benchmarks and much better coding and math, and moved every size except 72B to Apache 2.0.
• Qwen3 Alibaba Cloud · LLM · 2025 — Qwen3, the third generation of Alibaba's open (Apache 2.0) Qwen models: two mixture-of-experts models, the flagship Qwen3-235B-A22B (235B total, 22B active) and Qwen3-30B-A3B, plus six dense models from 0.6B to 32B. All switch between step-by-step thinking and fast non-thinking modes, were pretrained on about 36 trillion tokens in 119 languages and dialects, with better coding, agent and MCP support; Qwen says the flagship is competitive with DeepSeek-R1, o1, o3-mini, Grok-3 and Gemini 2.5 Pro.
• Qwen2.5 Alibaba Group · LLM · 2024 — Alibaba's Qwen2.5, released September 19, 2024 as dense open-weight base and instruct models in seven sizes from 0.5B to 72B parameters, pretrained on up to 18 trillion tokens, with more knowledge (MMLU 85+), stronger coding and math, better instruction following, long-text generation over 8K tokens and structured JSON output, a 128K context and support for over 29 languages; the specialized Qwen2.5-Coder and Qwen2.5-Math launched alongside, and all sizes except 3B and 72B are Apache 2.0.
• Allen AI FlexOlmo Allen AI · LLM · 2025 — Mixture-of-experts model (33B total parameters) built so data owners can contribute expert modules trained on their own data without sharing it, and switch that data's contribution on or off at inference. Apache 2.0.
• Allen AI Llama-3.1-Tulu-3-405B Allen AI · LLM · 2024 — Allen AI's instruction-tuned Llama 3.1 405B using Tulu 3 post-training recipe. Fully open training pipeline.
• Allen AI OLMo 2 32B Allen AI · LLM · 2025 — Fully open 32B model from Ai2's OLMo 2 family (Mar 2025), post-trained with Tülu 3 supervised fine-tuning, DPO and RLVR. Apache 2.0.
• ELMo Allen AI · Word Vectors · 2018 — Ai2's ELMo, deep contextualized word representations computed from the internal states of a deep bidirectional language model pretrained on a large text corpus, capturing syntax, semantics and how word meaning shifts with context (polysemy); adding ELMo to existing models significantly improved the state of the art on six NLP tasks, including question answering, textual entailment and sentiment analysis.
• OLMo Allen AI · LLM · 2024 — Fully open model including data and training code.
• Olmo 3 Allen AI · LLM · 2025 — Ai2's fully open 7B and 32B family (Base, Instruct, Think, RL-Zero) released with the entire model flow: data (Dolma 3, 9.3T tokens), code and every checkpoint. Olmo 3-Think 32B is billed as the first fully open 32B thinking model; OlmoTrace links outputs back to training data.
• SciBERT Allen AI · Science NLP · 2019 — Ai2's SciBERT, a BERT-based language model pretrained on a large multi-domain corpus of scientific publications to improve scientific NLP; Ai2 reported statistically significant gains over BERT and new state-of-the-art results on several tasks, including sequence tagging, sentence classification and dependency parsing, and released the code and models.
• Semantic Scholar Allen AI · Research AI · 2015 — AI-powered academic search. 200M+ papers.
• Tulu 3 Allen AI · LLM · 2024 — Ai2's Tülu 3, a family of fully open post-trained models built on Llama 3.1 base models and released with their data, code, training recipes and evaluation tools; training combines supervised fine-tuning, Direct Preference Optimization and a new method, Reinforcement Learning with Verifiable Rewards (RLVR), and Ai2 reports results above the instruct versions of Llama 3.1, Qwen 2.5 and Mistral and closed models such as GPT-4o-mini and Claude 3.5 Haiku.
• Tulu-3-405B Allen AI · LLM · 2025 — Ai2's Tülu 3 405B: Llama 3.1 405B post-trained with Ai2's fully open Tülu 3 recipe (supervised fine-tuning, DPO, then reinforcement learning with verifiable rewards), with its data, code and evaluation tools released; on Ai2's benchmark average it scored 80.7, above Llama 3.1 405B Instruct (79.0) and DeepSeek V3 (75.9) and below GPT-4o (81.6). Primarily English.
• WildGuard Allen AI · Safety Model · 2024 — Ai2's WildGuard, an open, lightweight moderation model (about 7B parameters, Apache 2.0) that in one tool flags harmful user prompts, detects unsafe model responses and judges whether a model refused, across 13 risk categories; trained on the 92K-example WildGuardMix data covering direct and adversarial jailbreak prompts, it beat ten open moderation models (up to 26.4% better on refusal detection), matched or beat GPT-4 in places, and cut jailbreak success in a chat interface from 79.8% to 2.4%.
• Alexa+ (LLM) Amazon · Virtual Assistant · 2025 — Alexa+, Amazon's next-generation Alexa assistant, introduced February 26, 2025 and built on large language models available on Amazon Bedrock; it orchestrates tens of thousands of services and devices through 'experts', can act on the web for users (for example arranging a repair through Thumbtack), remembers personal details, and works on Echo devices, Fire TV, the Alexa app and Alexa.com. It costs $19.99 a month or is free with Prime, and is available to all customers in the U.S. and Canada.
• AlexaTM Amazon · LLM · 2022 — Alexa Teacher Model: multi-billion-parameter encoders (up to 9.3B) pre-trained largely on spoken-form data and distilled into small models for Alexa's natural-language understanding.
• AlexaTM 20B Amazon · LLM · 2022 — 20B multilingual sequence-to-sequence model (Alexa Teacher Model) pre-trained on denoising and causal language modeling; Amazon showed it beats the much larger 540B PaLM on 1-shot summarization.
• Amazon CodeWhisperer Amazon · Code AI · 2023 — Amazon CodeWhisperer, AWS's real-time AI coding companion, generally available April 13, 2023 with a free Individual tier and a Professional tier; it suggested code in 15 languages across VS Code, JetBrains IDEs, AWS Cloud9 and the Lambda console, flagged suggestions resembling open-source training data with their repository URL and license, and scanned code for vulnerabilities. In its preview, users finished tasks 57% faster. On April 30, 2024 it became part of Amazon Q Developer.
• Amazon Nova Amazon · LLM · 2024 — Amazon Nova, Amazon's foundation model family on Amazon Bedrock, launched December 3, 2024 with Nova Micro (text only, lowest latency), Nova Lite (low-cost multimodal), Nova Pro (multimodal, balancing accuracy, speed and cost), Nova Canvas (images) and Nova Reel (video), later joined by Nova Premier and Nova Sonic; Amazon says they cover 200 languages and support fine-tuning, distillation and RAG. Nova 2 added Nova 2 Lite with extended thinking, Nova 2 Sonic and Nova Multimodal Embeddings.
• Amazon Nova 2 Lite Amazon · Reasoning LLM · 2025 — Fast, cost-effective reasoning model from the Nova 2 family (AWS re:Invent, Dec 2, 2025): processes text, images and video to generate text, with adjustable thinking depth; AWS positions it against Claude Haiku 4.5, GPT-5 mini and Gemini 2.5 Flash.
• Amazon Nova 2 Omni Amazon · Multimodal LLM · 2025 — Preview (Dec 2, 2025) of an all-in-one reasoning model that takes text, image, video and speech input and generates text and images; AWS calls it the first reasoning model of its kind. Covers multimodal understanding, natural-language image generation and editing, and speech transcription.
• Amazon Nova 2 Sonic Amazon · Voice / TTS · 2025 — Amazon's speech-to-speech model that unifies text and speech understanding and generation for real-time conversational AI, with expanded multilingual support and expressive voices (Nova 2 family, Dec 2, 2025).
• Amazon Nova Lite Amazon · Multimodal LLM · 2024 — Amazon Nova Lite, a very low-cost multimodal model from the first Amazon Nova generation, sold only through Amazon Bedrock, that reads text, images, video and documents (PDF, Word, Excel, CSV, HTML and more) and writes text, with a 300K-token context, up to 10K output tokens, support for 200+ languages (optimized for 15) and fine-tuning, aimed at chat, retrieval-augmented generation, agents and video analysis; Amazon has since released Nova 2 models.
• Amazon Nova Micro Amazon · LLM · 2024 — Amazon Nova Micro, the text-only model of the first Amazon Nova generation, sold only through Amazon Bedrock and built for the lowest latency at very low cost, with a 128K-token context, up to 10K output tokens, support for 200+ languages (optimized for 15) and fine-tuning, aimed at chat, retrieval-augmented generation and agent applications; Amazon has since released Nova 2 models.
• Amazon Nova Pro Amazon · LLM · 2024 — Amazon Nova Pro (December 3, 2024), a multimodal model in Amazon's Nova family on Amazon Bedrock with what Amazon calls the best mix of accuracy, speed and cost for a wide range of tasks: it takes text, image, video and document inputs with a 300K-token context and up to 10K output tokens, supports 200+ languages and fine-tuning, and suits chat, RAG, agents, video analysis and UI automation; Amazon says it costs at least 75% less than the best-performing models in its class on Bedrock.
• Amazon Q Developer Amazon · Code AI · 2024 — Amazon Q Developer (April 30, 2024), AWS's generative AI assistant for software development, which absorbed Amazon CodeWhisperer: it offers code suggestions, reviews and agentic tasks (implementing features, documenting, testing, refactoring, upgrades) in JetBrains, VS Code, Visual Studio, Eclipse and the command line, answers AWS questions in the console, Teams and Slack, and ports .NET to Linux and upgrades Java. AWS ends support for its IDE plugins on April 30, 2027 and points users to Kiro.
• Amazon Titan Embeddings Amazon · Embeddings · 2023 — Amazon's Titan Embeddings models on Amazon Bedrock: Titan Text Embeddings V2 turns up to 8,192 tokens (50,000 characters) into 1,024-, 512- or 256-dimension vectors for retrieval, semantic similarity and clustering, optimized for English with 100+ languages in preview, alongside the earlier Titan Text Embeddings G1 and Titan Multimodal Embeddings G1, which maps text, images or both into one vector space (1,024, 384 or 256 dimensions) for multimodal search and recommendations.
• Amazon Titan Image Generator Amazon · Image Generation · 2023 — Amazon's Titan Image Generator, a text-to-image model sold through Amazon Bedrock that also edited images (inpainting, outpainting, mask-free text-guided edits and variations) and transferred styles from reference images; G1 v2 added layout guidance from reference images, background removal, color-palette control and subject-consistency fine-tuning, and every image carried an invisible watermark and C2PA metadata. Bedrock listed June 30, 2026 as its end-of-life date.
• Amazon Titan Text Amazon · LLM · 2023 — Amazon Titan Text, AWS's text-generation foundation models for Amazon Bedrock, previewed April 13, 2023 for summarization, text generation, classification, Q&A and information extraction: Titan Text Express (8K-token context, English-optimized, 100+ languages in preview) and Titan Text Lite (4K context, fastest, built for fine-tuning) became generally available November 29, 2023, and Titan Text Premier, tuned for RAG and agents, followed May 7, 2024. Bedrock's Titan list no longer includes them.
• AMD Instella-3B AMD · LLM · 2025 — AMD Instella is a fully open 3B language-model family released with weights, code, data and training details for reproducible LLM research.
• Ling-3.0-Flash Ant Group · LLM · 2026 — Ant Group inclusionAI efficiency MoE: 124B parameters activating only ~5.1B per token, which Ant says matches or beats its own trillion-parameter flagship on most published benchmarks. Native hybrid-linear attention plus cluster-level hierarchical caching (claimed 60-80% time-to-first-token cut on long inputs).
• Ling-3.0-Flash-Fin Ant Group · LLM · 2026 — Finance fine-tune of Ling-3.0-Flash: 124B total / 5.1B active, 262K context. MIT-licensed weights published (unlike the Sante health fine-tune).
• Ling-3.0-Flash-Sante Ant Group · LLM · 2026 — Health/medicine fine-tune of Ling-3.0-Flash: 124B total / 5.1B active, 262K context. API-only, no open weights.
• Claude Anthropic · LLM · 2023 — Anthropic's first publicly available Claude models, launched March 14, 2023 after a closed alpha with partners such as Notion, Quora and DuckDuckGo: a next-generation AI assistant built on Anthropic's research into helpful, honest and harmless AI, offered through a chat interface and the API in two versions, Claude and the lighter, cheaper and faster Claude Instant; a May 2023 update expanded its context window from 9K to 100K tokens. The Claude 1 models were retired November 6, 2024.
• Claude 2 Anthropic · LLM · 2023 — Anthropic's Claude 2, released July 11, 2023 through the API and the new public beta of claude.ai in the US and UK, with a 100K-token context window (hundreds of pages, or even a book) and longer outputs; it scored 76.5% on the multiple-choice Bar exam (up from 73.0% for Claude 1.3), 71.2% on the Codex HumanEval coding test (from 56.0%) and 88.0% on GSM8K math, and was twice as good at giving harmless responses in Anthropic's red-teaming evaluation. Retired July 21, 2025.
• Claude 2.1 Anthropic · LLM · 2023 — Anthropic's Claude 2.1 (November 21, 2023), which doubled Claude's context window to an industry-leading 200K tokens (about 150,000 words, or over 500 pages), halved false statements compared with Claude 2.0, cut incorrect answers on long documents by 30%, and added system prompts and a beta of tool use for calling developer-defined functions and APIs; it powered claude.ai and the API. Retired July 21, 2025.
• Claude 3 Anthropic · Multimodal LLM · 2024 — Anthropic's Claude 3 family, announced March 4, 2024 in three tiers, Haiku, Sonnet and Opus, all with image understanding (photos, charts, graphs and technical diagrams), a 200K-token context window, fewer unnecessary refusals and better non-English output; Opus led most common benchmarks at launch, and the release made the Claude API generally available in 159 countries. The Claude 3 models were retired between July 2025 and April 2026.
• Claude 3 Haiku Anthropic · LLM · 2024 — Anthropic's Claude 3 Haiku, the fastest and most compact Claude 3 model, announced March 4, 2024 and released March 13 at $0.25/$1.25 per million input/output tokens; it processed about 21K tokens (roughly 30 pages) per second on prompts under 32K tokens, three times faster than its peers, with vision and a 200K-token context, so that $1 could analyze about 400 Supreme Court cases or 2,500 images. Retired April 20, 2026.
• Claude 3 Opus Anthropic · LLM · 2024 — Anthropic's Claude 3 Opus (March 4, 2024), the most intelligent Claude 3 model, which outperformed peers on most common benchmarks including MMLU, GPQA and GSM8K, with near-human comprehension and fluency on complex tasks, vision, a 200K-token context (1M for select customers) and over 99% recall in needle-in-a-haystack tests; it cost $15/$75 per million input/output tokens and came to claude.ai for Pro subscribers. Retired January 5, 2026.
• Claude 3 Sonnet Anthropic · LLM · 2024 — Anthropic's Claude 3 Sonnet (March 4, 2024), the middle Claude 3 model, balancing intelligence and speed for enterprise workloads at $3/$15 per million input/output tokens; it was twice as fast as Claude 2 and 2.1 with higher intelligence, read images and had a 200K-token context, powered the free claude.ai experience at launch and was the first Claude 3 model on Amazon Bedrock. Retired July 21, 2025.
• Claude 3.5 Anthropic · Multimodal LLM · 2024 — Anthropic's Claude 3.5 family: Claude 3.5 Sonnet (June 2024) beat Claude 3 Opus at twice its speed and Sonnet's price and launched alongside Artifacts on claude.ai; an upgraded 3.5 Sonnet (October 22, 2024) raised SWE-bench Verified from 33.4% to 49.0% and introduced computer use in public beta; and Claude 3.5 Haiku matched or beat Claude 3 Opus on many benchmarks at Haiku speed. 3.5 Sonnet was retired October 28, 2025 and 3.5 Haiku February 19, 2026.
• Claude 3.5 Haiku Anthropic · LLM · 2024 — Anthropic's Claude 3.5 Haiku, announced October 22, 2024 as the next generation of its fastest model: at a similar speed to Claude 3 Haiku it improved across every skill and surpassed Claude 3 Opus on many intelligence benchmarks, scoring 40.6% on SWE-bench Verified, ahead of many agents built on the original Claude 3.5 Sonnet and GPT-4o; it launched text-only, with image input to follow, for user-facing products and sub-agent tasks. Retired February 19, 2026.
• Claude 3.5 Sonnet Anthropic · Multimodal LLM · 2024 — Anthropic's Claude 3.5 Sonnet (June 2024), the first Claude 3.5 model, which outperformed Claude 3 Opus and set new marks on GPQA, MMLU and HumanEval at twice Opus's speed and Claude 3 Sonnet's price ($3/$15 per million tokens, 200K context); it solved 64% of problems in Anthropic's agentic coding evaluation against 38% for Opus, was its strongest vision model yet, and launched free on claude.ai alongside Artifacts. Retired October 28, 2025.
• Claude 3.5 Sonnet (Oct 2024) Anthropic · Multimodal LLM · 2024 — Anthropic's upgraded Claude 3.5 Sonnet (October 22, 2024), with across-the-board gains over the June version at the same price and speed, especially in agentic coding and tool use: SWE-bench Verified rose from 33.4% to 49.0%, ahead of all public models at the time, and TAU-bench retail from 62.6% to 69.2%; it was the first frontier model to offer computer use in public beta, scoring 14.9% on OSWorld's screenshot-only tests against 7.8% for the next best. Retired October 28, 2025.
• Claude 3.7 Anthropic · Multimodal LLM · 2025 — Anthropic's Claude 3.7 generation, consisting of Claude 3.7 Sonnet (February 24, 2025), which Anthropic called the first hybrid reasoning model on the market: one model that answers near-instantly or thinks step by step in visible extended thinking, with API users able to cap its thinking budget at up to 128K tokens; it brought strong coding and front-end gains, launched alongside the Claude Code research preview at $3/$15 per million tokens, and was retired February 19, 2026.
• Claude 3.7 Sonnet Anthropic · Multimodal LLM · 2025 — Anthropic's Claude 3.7 Sonnet (February 24, 2025), billed as the first hybrid reasoning model on the market: it answers near-instantly or uses visible, step-by-step extended thinking, with API users able to set a thinking budget up to its 128K-token output limit; state of the art on SWE-bench Verified and TAU-bench at launch and strong at front-end web development, it was on all Claude plans (extended thinking for paid users) at $3/$15 per million tokens. Retired February 19, 2026.
• Claude 4 Anthropic · Multimodal LLM · 2025 — Anthropic's Claude 4 generation, launched May 22, 2025 with Claude Opus 4 (which Anthropic called the world's best coding model, at 72.5% on SWE-bench and able to work for hours) and Claude Sonnet 4 (72.7% on SWE-bench), both hybrid models with near-instant answers or extended thinking that can now use tools while thinking; the launch also made Claude Code generally available and added API code execution, an MCP connector and a Files API. Both models were retired June 15, 2026.
• Claude Code Anthropic · Code Agent · 2025 — Anthropic's Claude Code, an agentic coding tool that reads a codebase, plans changes, runs tests and carries out fixes or multi-day migrations, steered from the terminal, VS Code and JetBrains IDEs, the desktop and mobile apps, the web or Slack; it began as a research preview with Claude 3.7 Sonnet in February 2025, became generally available with Claude 4 in May 2025, and is included in paid Claude plans or billed at API rates through a Console account.
• Claude Fable 5 Anthropic · Multimodal LLM · 2026 — Claude Fable 5 (June 9, 2026), Anthropic's first Mythos-class model made safe for general use: it shares its underlying model with Claude Mythos 5 but adds safeguards that send some cybersecurity and other risky requests to Claude Opus 4.8. Anthropic reports 80.3% on SWE-Bench Pro and 85.0% on OSWorld-Verified, with gains in long-horizon coding, knowledge work and vision. A US export-control directive suspended access from June 12 to July 1, 2026; Fable 5.1 has since succeeded it.
• Claude Fable 5.1 Anthropic · LLM · 2026 — Incremental update to Anthropic's frontier flagship Claude Fable 5: 1M-token context, 128K max output, and a cache-read price cut to $0.25/M. Artificial Analysis Intelligence Index 66; Terminal-Bench-Science 52.6%.
• Claude Haiku 4.5 Anthropic · LLM · 2025 — Anthropic's Claude Haiku 4.5 (October 15, 2025), a small model with coding performance similar to the earlier frontier model Claude Sonnet 4 at one-third the cost and more than twice the speed, and better computer use; built for real-time, low-latency work such as chat assistants, customer service agents, pair programming and sub-agents orchestrated by Sonnet 4.5, it costs $1/$5 per million tokens (API ID claude-haiku-4-5-20251001) and has since been followed by Claude Haiku 5.5.
• Claude in Chrome Anthropic · Browser Agent · 2025 — Anthropic's Claude in Chrome, a Chrome extension that lets Claude see the page you're on, click buttons, fill in forms and navigate sites on your behalf; piloted from August 2025 as a research preview with 1,000 Max subscribers while Anthropic built defenses against prompt injection, it opened in beta to all Max users in November 2025 (adding scheduled tasks and multi-tab workflows) and to Pro, Team and Enterprise plans in December 2025, with a Claude Code integration and admin site controls.
• Claude Instant Anthropic · LLM · 2023 — Anthropic's Claude Instant, the lighter, less expensive and much faster sibling of Claude launched March 14, 2023 for tasks such as casual dialogue, text analysis, summarization and document comprehension; version 1.2 (August 2023) drew on Claude 2's strengths, raising Codex coding scores from 52.8% to 58.7% and GSM8K math from 80.9% to 86.7%, with fewer hallucinations. All Claude Instant models were retired November 6, 2024.
• Claude Mythos 5 Anthropic · Multimodal LLM · 2026 — Claude Mythos 5 (June 9, 2026), the same underlying model as Claude Fable 5 with some safeguards lifted, offered to a small group of cyber defenders through Project Glasswing, replacing Claude Mythos Preview; Anthropic calls its cybersecurity capabilities the strongest of any model and says it sped parts of drug design tenfold. A US export-control directive suspended it on June 12, 2026; access returned for approved US organizations after a June 26 government approval. Mythos 5.1 followed later.
• Claude Mythos 5.1 Anthropic · LLM · 2026 — Same weights as Claude Fable 5.1 but with some safeguards lifted for approved frontier red-team and enterprise partners under Project Glasswing. Terminal-Bench 4.0: 60.9%.
• Claude Mythos Preview Anthropic · LLM · 2026 — Claude Mythos Preview, Anthropic's first Mythos-class model, announced April 7, 2026 as a gated research preview for Project Glasswing, a defensive-cybersecurity effort whose launch partners include AWS, Apple, Cisco, CrowdStrike, Google, Microsoft and NVIDIA, plus 40+ other organizations; Anthropic says it found thousands of high-severity vulnerabilities, including in every major operating system and web browser. Claude Mythos 5 replaced it in June 2026, and Anthropic lists it as deprecated.
• Claude Opus 4 Anthropic · Multimodal LLM · 2025 — Anthropic's Claude Opus 4 (May 22, 2025), which Anthropic called the world's best coding model at launch, leading SWE-bench (72.5%) and Terminal-bench (43.2%) and able to work continuously for several hours on long agentic tasks; a hybrid model with near-instant answers or extended thinking with tool use, it kept 'memory files' when given local file access, cost $15/$75 per million tokens and launched with safety measures for higher AI Safety Levels such as ASL-3. Retired June 15, 2026.
• Claude Opus 4.1 Anthropic · Multimodal LLM · 2025 — Anthropic's Claude Opus 4.1 (August 5, 2025), an upgrade to Claude Opus 4 on agentic tasks, real-world coding and reasoning that raised SWE-bench Verified to 74.5% and improved in-depth research and data analysis, especially detail tracking and agentic search; GitHub, Rakuten and Windsurf reported gains in multi-file refactoring and precise debugging. It kept Opus 4's pricing and came to paid Claude users, Claude Code and the API. Retired August 5, 2026.
• Claude Opus 4.5 Anthropic · Multimodal LLM · 2025 — Anthropic's Claude Opus 4.5 (November 24, 2025), which Anthropic called the best model in the world for coding, agents and computer use, and stronger at deep research, slides and spreadsheets; it cut Opus pricing to $5/$25 per million tokens and added an effort control (at medium effort it matched Sonnet 4.5's best SWE-bench Verified score with 76% fewer output tokens), and within a 2-hour limit it outscored every human candidate on Anthropic's performance-engineering take-home exam.
• Claude Opus 4.6 Anthropic · LLM · 2026 — Anthropic's Claude Opus 4.6 (February 5, 2026), an upgrade to Opus 4.5 that plans more carefully, sustains agentic tasks longer and reviews and debugs its own code better, and the first Opus model with a 1M-token context window (in beta); it topped Terminal-Bench 2.0, Humanity's Last Exam and BrowseComp at launch and beat GPT-5.2 by about 144 Elo on GDPval-AA, and arrived with agent teams in Claude Code, adaptive thinking and effort controls, at $5/$25 per million tokens.
• Claude Opus 4.7 Anthropic · LLM · 2026 — Anthropic's Opus model announced Apr 16, 2026, with major gains on the hardest coding tasks, long-running agents and high-resolution vision; since superseded by Anthropic's Claude 5-series models.
• Claude Opus 4.8 Anthropic · Multimodal LLM · 2026 — Flagship Opus revision, 41 days after 4.7. Dynamic Workflows in Claude Code (hundreds of parallel subagents), user-selectable effort control, Fast mode tier. ~4x less likely than 4.7 to let flaws in its own code pass unflagged; 1M-token context. Anthropic framed it as "a modest but tangible improvement."
• Claude Opus 5 Anthropic · LLM · 2026 — Fourth Claude 5-series model in under two months (after Sonnet 5, Fable 5, Mythos 5). Five-level effort toggle (low/medium/high/xhigh/max), 1M-token context. Leads Frontier-Bench v0.1 (more than doubling Opus 4.8 at lower cost per task); within 0.5% of Fable 5's CursorBench 3.2 peak at half the per-task cost; 3x next-best on ARC-AGI 3.
• Claude Opus 5.5 Anthropic · LLM · 2026 — First model in Anthropic's Claude 5.5 family. Anthropic says it performs at Claude Fable 5.1 level on most work while costing about 40% less than Opus 5 on typical workloads, with >30% faster output. Launch results: Terminal-Bench 4.0 66.4%, GDPval-AA v2.1 1846 Elo, OSWorld 2.0 81.8% (partial), Humanity's Last Exam 67.7% (with tools). API model claude-opus-5-5; thinking cannot be disabled; zero data retention available.
• Claude Projects Anthropic · AI Workspace · 2024 — Claude Projects, launched June 25, 2024 for Claude.ai Pro and Team users, let people group chats around a curated knowledge base of documents, code and notes (a 200K-token context, about a 500-page book) with project-level custom instructions, and let Team users share chats to a project activity feed; today Claude's free plan allows up to five projects, paid plans include them, and Team and Enterprise plans add project sharing and collaboration.
• Claude Sonnet 4 Anthropic · LLM · 2025 — Anthropic's Claude Sonnet 4 (May 22, 2025), a significant upgrade to Claude 3.7 Sonnet with state-of-the-art coding (72.7% on SWE-bench), more precise instruction following and 65% less shortcut-taking on agentic tasks; a hybrid model with near-instant answers or extended thinking, it was open to free Claude users, cost $3/$15 per million API tokens and was picked by GitHub to power the new coding agent in GitHub Copilot. Retired June 15, 2026.
• Claude Sonnet 4.5 Anthropic · Multimodal LLM · 2025 — Anthropic's Claude Sonnet 4.5 (September 29, 2025), which Anthropic called the best coding model in the world and the best at using computers at launch: state of the art on SWE-bench Verified, able to stay on complex multi-step tasks for more than 30 hours, and leading OSWorld at 61.4% (up from Sonnet 4's 42.2%); it shipped with Claude Code checkpoints, a VS Code extension, the Claude Agent SDK and API memory and context-editing tools, at $3/$15 per million tokens. Retiring November 30, 2026.
• Claude Sonnet 4.6 Anthropic · Multimodal LLM · 2026 — Anthropic's Claude Sonnet 4.6 (February 17, 2026), its most capable Sonnet at the time and a full upgrade across coding, computer use, long-context reasoning, agent planning, knowledge work and design, with a 1M-token context window in beta; it became the default model for Free and Pro users in claude.ai and Claude Cowork at $3/$15 per million tokens, early-access developers often preferred it even to Opus 4.5, and Claude Code testers chose it over Sonnet 4.5 about 70% of the time.
• Claude Sonnet 5 Anthropic · LLM · 2026 — Anthropic's most agentic Sonnet: plans, uses tools like browsers and terminals, and runs autonomously at a level that previously required larger models. Near Opus 4.8 on agentic coding and computer use; outscores it on knowledge work. New tokenizer encodes ~1.0-1.35x more tokens than Sonnet 4.6 (effective cost note). Model string claude-sonnet-5.
• Claude.ai Anthropic · Chat Interface · 2023 — Anthropic's Claude app, opened to the public as a beta at claude.ai in the US and UK on July 11, 2023 alongside Claude 2 and now on web, desktop and mobile; the free plan covers chat, web search, file creation, code execution, memory, app connectors and Artifacts, Pro ($20 a month, or $17 billed annually) adds more usage, Claude Code, task scheduling, more models and Claude in Chrome, and Max (from $100 a month) offers 5x or 20x Pro's usage.
• Cursor Anysphere · Code Assistant / IDE · 2023 — Cursor, Anysphere's AI coding agent and editor, acquired by SpaceX in August 2026: agents plan, write and test code in the desktop app, run in parallel on their own cloud computers to build and demo features end to end, and also work from the terminal, in Slack and in GitHub pull-request reviews (Bugbot), using SpaceXSI's Grok models and those of other labs. Since October 2026 it is sold through The X Subscription, from a free tier to Ultra at $200 a month, with Teams from $30 per user.
• Apple DiffuCoder Apple · Code LLM · 2025 — Apple's 7B masked diffusion language model for code, which generates code by iteratively denoising a sequence rather than writing left to right; the Instruct version is tuned from DiffuCoder-7B-Base on OpenCoder-SFT data. Weights under Apple's research-only license.
• Apple Foundation Model (AFM) Apple · LLM · 2024 — Apple's foundation language models behind Apple Intelligence: a roughly 3-billion-parameter model built to run efficiently on device and a larger server model that runs on Private Cloud Compute; Apple's report covers their architecture, training data, training process, inference optimization, evaluation and Responsible AI principles.
• Apple Intelligence Apple · AI Platform · 2024 — System-wide AI features across Apple devices.
• Apple MLX Apple · ML Framework · 2023 — Apple's MLX, an array framework for machine learning on Apple silicon from Apple machine learning research, with a NumPy-like Python API plus C++, C and Swift APIs, PyTorch-like mlx.nn and mlx.optimizers packages, composable function transformations (automatic differentiation, vectorization, graph optimization), lazy computation, dynamic graphs and a unified memory model in which arrays live in shared memory across CPU and GPU.
• Ferret Apple · Multimodal LLM · 2023 — Apple's Ferret (ICLR 2024 Spotlight), an end-to-end multimodal LLM built on Vicuna that accepts any-form referring and grounds anything in its responses, using a hybrid region representation and a spatial-aware visual sampler; Apple released 7B and 13B weight deltas over Vicuna (research-only, CC BY-NC), the 1.1M-sample GRIT instruction dataset and Ferret-Bench, followed by Ferret-v2 and Ferret-UI.
• OpenELM Apple · LLM · 2024 — Apple's OpenELM family of small open language models (270M, 450M, 1.1B and 3B parameters, each pretrained and instruction-tuned), which use layer-wise scaling to allocate parameters within each transformer layer; pretrained on about 1.8 trillion tokens from RefinedWeb, deduplicated PILE and subsets of RedPajama and Dolma.
• Siri (Apple Intelligence) Apple · Virtual Assistant · 2024 — Siri powered by Apple Foundation Models.
• Arc Institute Evo 2 Arc Institute · Science/Biology · 2025 — Arc Institute Evo 2 is an AI model or product entry requiring row-level review for exact capabilities and positioning.
• Artbreeder Artbreeder · Image Generation · 2019 — Collaborative GAN-based image mixing.
• Asana Intelligence Asana · Productivity AI · 2023 — AI project management features.
• BGE BAAI · Embeddings · 2023 — BAAI's BGE (BAAI General Embedding), an open family of embedding and reranking models for search and retrieval-augmented generation: the first bge-large models (August 2, 2023) ranked first on the MTEB and C-MTEB benchmarks, followed by English and Chinese v1.5 embedders in large, base and small sizes, the multilingual BGE-M3 with dense, sparse and multi-vector retrieval over inputs up to 8,192 tokens, LLM-based embedders and cross-encoder rerankers. MIT license.
• BGE-Large BAAI · Embeddings · 2023 — BAAI's English BGE large text embedding model for retrieval, classification, clustering and semantic search; BAAI now recommends the newer bge-large-en-v1.5, which has a more reasonable similarity distribution and is used the same way.
• BGE-M3 BAAI · Embeddings · 2024 — BAAI's BGE-M3 embedding model, which performs dense, multi-vector and sparse retrieval in one model, works in more than 100 languages and handles inputs from short sentences up to 8,192 tokens.
• EVA-02 BAAI · Vision/CV · 2023 — BAAI's EVA-02 (March 2023), a family of plain Transformer image encoders from 6M to 304M parameters pre-trained by masked image modeling to reconstruct language-aligned features from a giant EVA-CLIP teacher; using only public data, the 304M model reaches 90.0% top-1 accuracy on ImageNet-1K after fine-tuning, and its EVA-02-CLIP variant reaches 80.4% zero-shot, beating the largest open CLIP of the time with about a sixth of the parameters and training data.
• Baichuan 2 Baichuan · LLM · 2023 — Baichuan's Baichuan 2 open language models, trained on a 2.6-trillion-token corpus and released in 7B and 13B Base and Chat versions plus 4-bit quantized Chat models; Baichuan says they set the best results among same-size models on major Chinese and English benchmarks, and a Dec 29, 2023 update to the 13B Chat model (v2) improved math, logical reasoning and complex instruction following.
• Baichuan-M1-14B Baichuan · LLM · 2025 — Open-source 14B LLM trained from scratch by Baichuan and optimized for medical scenarios.
• Ernie 3.0 Titan Baidu · LLM · 2021 — Baidu's ERNIE 3.0 Titan, a knowledge-enhanced Chinese pretrained language model with up to 260 billion parameters, trained on the PaddlePaddle platform by scaling up the 10B ERNIE 3.0; it adds self-supervised adversarial and controllable language-modeling losses for credible, controllable text plus an online distillation framework, and Baidu reported state-of-the-art results on 68 NLP datasets, calling it the largest Chinese dense pretrained model at the time.
• ERNIE 4.0 Baidu · Multimodal LLM · 2023 — Baidu's ERNIE 4.0 large language model; since succeeded by the ERNIE 4.5 family, which Baidu has released as open weights on Hugging Face.
• ERNIE 5 Baidu · LLM · 2025 — Baidu's natively omni-modal foundation model, unveiled at Baidu World 2025; jointly models text, images, audio and video for multimodal understanding and generation. Proprietary.
• ERNIE Bot Baidu · Chat Interface · 2023 — Baidu's consumer AI chatbot.
• Ernie Bot (Wenxin Yiyan) Baidu · Chat Interface · 2023 — Chinese name for ERNIE Bot consumer product.
• ERNIE X1.1 Baidu · LLM · 2025 — Baidu reasoning model unveiled at WAVE SUMMIT 2025; Baidu reports gains in factuality, instruction following and agentic capability, surpassing DeepSeek-R1-0528 overall and on par with GPT-5 and Gemini 2.5 Pro (vendor claims).
• ERNIE-4.5-21B-A3B-Thinking Baidu · Reasoning LLM · 2025 — Open-sourced by Baidu at WAVE SUMMIT 2025: a 21B-total / 3B-active MoE thinking (reasoning) variant of ERNIE 4.5.
• Salamandra Barcelona Supercomputing Center · LLM · 2024 — Barcelona Supercomputing Center Salamandra. Barcelona Supercomputing Center LLM model.
• Base 1 Base44 · Code LLM · 2026 — First in-house LLM from a vibe-coding platform: proprietary model trained on tens of millions of real app-building interactions; auto-routing sends tasks to Base 1 when it beats alternatives on internal benchmarks.
• Beatoven.ai Beatoven · Music Generation · 2022 — AI background music for content.
• Beautiful.ai Beautiful.ai · Presentation AI · 2023 — AI-powered presentation design.
• EVA-01 Beijing Academy of AI · Vision/CV · 2023 — EVA-01. Beijing Academy of AI Vision/CV model.
• Remini Bending Spoons · Image Enhancement · 2023 — AI photo enhancer and unblurring.
• Berkeley Sky-T1-32B-Preview Berkeley · LLM · 2025 — 32B reasoning model fine-tuned from Qwen2.5-32B-Instruct on 17K examples by UC Berkeley's NovaSky team; on par with o1-preview in math and coding. Apache 2.0.
• StarCoder 2 BigCode (Hugging Face) · Code LLM · 2024 — 15B code model. 600+ languages. 16K context. BigCode Open RAIL-M license.
• SantaCoder BigCode/Hugging Face · Code LLM · 2023 — BigCode's 1.1B-parameter SantaCoder code models, trained with Fill-in-the-Middle on the Python, Java and JavaScript subset of The Stack v1.1 (opt-outs excluded); the main model uses multi-query attention, a 2,048-token context and 236B training tokens. A base model, not instruction-tuned.
• StarCoder BigCode/Hugging Face · Code LLM · 2023 — BigCode's StarCoder, a 15.5B-parameter open code model with an 8K context, fill-in-the-middle infilling and multi-query attention for fast inference; its base, StarCoderBase, was trained on 1 trillion tokens from The Stack (permissively licensed GitHub code in more than 80 programming languages plus GitHub issues, commits and notebooks), and StarCoder adds 35B Python tokens, reaching 40% pass@1 on HumanEval with prompting. Released under the BigCode OpenRAIL-M license.
• StarCoder2-15B BigCode/Hugging Face · Code LLM · 2024 — BigCode's 15B-parameter StarCoder2 code model, trained with the Fill-in-the-Middle objective on 4+ trillion tokens from 600+ programming languages in The Stack v2 (opt-out requests excluded), with grouped-query attention and a 16,384-token context (4,096-token sliding window); a base model, not instruction-tuned.
• StarCoder2-3B BigCode/Hugging Face · Code LLM · 2024 — BigCode's 3B-parameter StarCoder2 code model, trained with the Fill-in-the-Middle objective on 3+ trillion tokens from 17 programming languages in The Stack v2 (opt-out requests excluded), with grouped-query attention and a 16,384-token context (4,096-token sliding window); a base model, not instruction-tuned.
• StarCoder2-7B BigCode/Hugging Face · Code LLM · 2024 — BigCode's 7B-parameter StarCoder2 code model, trained with the Fill-in-the-Middle objective on 3.5+ trillion tokens from 17 programming languages in The Stack v2 (opt-out requests excluded), with grouped-query attention and a 16,384-token context (4,096-token sliding window); a base model, not instruction-tuned.
• FLUX 1.1 Pro Black Forest Labs · Image Generation · 2024 — Black Forest Labs' FLUX1.1 [pro] (October 2024), a text-to-image model that BFL says generates six times faster than FLUX.1 [pro] with better image quality, prompt adherence and diversity; tested as 'blueberry', it took the highest Elo in the Artificial Analysis image arena (as of October 1, 2024). It launched with the beta BFL API at $0.04 per image and through Together.ai, Replicate, fal.ai and Freepik. BFL's team created VQGAN, Latent Diffusion and Stable Diffusion models.
• FLUX 1.1 Ultra Black Forest Labs · Image Generation · 2024 — Black Forest Labs' FLUX1.1 [pro] Ultra and Raw modes (November 6, 2024): Ultra generates images at four times standard FLUX1.1 [pro] resolution, up to 4 megapixels, in about 10 seconds per sample without losing prompt adherence, which BFL says is over 2.5x faster than comparable high-resolution offerings, at $0.06 per image; Raw mode gives a less synthetic, candid-photography look, with more diverse human subjects and more realistic nature photos.
• FLUX 3 Black Forest Labs · Video Generation · 2026 — FLUX 3 (July 23, 2026), Black Forest Labs' first multimodal foundation model, trained jointly on images, video and audio: one model makes videos up to 20 seconds long with native dialogue, sound effects and ambience from text, images, keyframes or a clip to continue, with multilingual lip-sync and multi-shot scenes, and also synthesizes images and predicts robot actions (FLUX-mimic). FLUX 3 Video became generally available on August 4, 2026; FLUX 3 Image and open FLUX 3 Dev come later.
• FLUX.1 Dev Black Forest Labs · Image Generation · 2024 — Black Forest Labs' FLUX.1 [dev] (August 1, 2024), an open-weight 12B-parameter rectified flow transformer for text-to-image, guidance-distilled directly from FLUX.1 [pro] to reach similar quality and prompt adherence more efficiently; BFL said FLUX.1 [pro] and [dev] beat Midjourney v6.0, DALL-E 3 (HD) and SD3-Ultra on visual quality, prompt following, size and aspect variability, typography and output diversity. Weights carry BFL's non-commercial license, but outputs may be used commercially.
• FLUX.1 Kontext Black Forest Labs · Image Editing · 2025 — Flow-matching model suite that unifies image generation with instruction-based editing from text and reference images (character consistency, local edits). [pro] and [max] shipped via API on May 29, 2025; an open 12B [dev] checkpoint followed in June 2025.
• FLUX.1 Pro Black Forest Labs · Image Generation · 2024 — Black Forest Labs' original 12B FLUX.1 model for API use (August 2024), since followed by FLUX 1.1 [pro], FLUX.2 and FLUX 3; still on BFL's API price list.
• FLUX.1 Schnell Black Forest Labs · Image Generation · 2024 — Black Forest Labs' FLUX.1 [schnell] (August 1, 2024), the fastest of the original FLUX.1 models: a 12B-parameter rectified flow transformer for text-to-image, built from hybrid multimodal and parallel diffusion transformer blocks and trained with latent adversarial diffusion distillation to make high-quality images in 1 to 4 steps. It is released under Apache 2.0 for personal, scientific and commercial use, and BFL said it outperformed Midjourney v6.0 and DALL-E 3 (HD).
• FLUX.2 Black Forest Labs · Image Generation · 2025 — BFL's second-generation family: [pro] and [flex] API tiers plus an open 32B [dev] checkpoint, with one model handling both creation and editing; later extended by FLUX.2 [max] (Dec 16, 2025) and [klein] 4B/9B (Jan 15, 2026).
• Bland Speech v3 Bland AI · Voice / TTS · 2026 — Tops Design Arena's blind pairwise Audio Realism benchmark at 1365 Elo — above ElevenLabs, Microsoft MAI-Voice-2 and Grok TTS, second only to real human recordings (~1500). Trained on 100M+ real phone conversations; keeps the breaths, hesitations and fillers TTS usually sands off.
• BloombergGPT Bloomberg L.P. · LLM · 2023 — 50B finance-domain LLM. Trained on Bloomberg financial data. Internal research model. March 2023.
• Boomy Boomy · Music Generation · 2019 — Create and monetize AI music.
• Brave Leo Brave · Browser AI · 2023 — AI assistant in Brave browser.
• Brave Search AI Brave · Search AI · 2023 — AI-powered search answers.
• Coze ByteDance · AI Platform · 2024 — ByteDance bot-building platform.
• Doubao (ByteDance) ByteDance · Chat Interface · 2024 — ByteDance's consumer AI chatbot. 600M MAU.
• Doubao-1.5-pro ByteDance · Multimodal LLM · 2025 — ByteDance's mixture-of-experts flagship (Jan 2025) built with training-inference co-design; ByteDance says it outperforms top dense pre-trained models while activating far fewer parameters.
• Seed-Thinking-v1.5 ByteDance · Reasoning LLM · 2025 — ByteDance Seed reasoning model that thinks before responding: a mixture-of-experts with 20B activated and 200B total parameters, scoring 86.7 on AIME 2024, 55.0 on Codeforces and 77.3 on GPQA.
• Seedance ByteDance · Video Generation · 2025 — Up to 12s clips. Strong motion coherence. Pairs with Seedream for stills.
• Seedance 2.5 ByteDance · Video Generation · 2026 — ByteDance's top-ranked video model goes global: native 30-second clips, long-video mode assembling up to 3 minutes with consistent characters, up to 50 multimodal references per generation (incl. 3D white models and green-screen footage), timestamp control to the second, and Maya/Blender plugins. #1 Arena video ranking at launch.
• SeedRealtime ByteDance · Speech AI · 2026 — Native audio-visual full-duplex LLM replacing cascaded ASR-VLM-TTS pipelines: a single end-to-end model that listens, watches and speaks simultaneously, with turn-taking inside the model and no external VAD; 50% fewer conversational pacing failures in human evals; ties voices to faces in noisy rooms and speaks up proactively on scene changes.
• Seedream 5.0 Pro ByteDance · Image Generation · 2026 — Flagship tier of the Seedream 5 line, pitched as a shift from image generator to design tool: interactive precision editing (point, lasso, sketch), intelligent layer separation into editable layers, dense infographic rendering, native text in 10+ languages, 4K max resolution.
• UI-TARS-72B ByteDance · LLM · 2025 — Native GUI agent model that operates graphical interfaces with human-like perception, reasoning and action, combining perception, grounding and memory in one vision-language model. Apache 2.0.
• Mars5 Camb.AI · Speech Synthesis · 2024 — CAMB.AI's MARS5, an open English text-to-speech model built for hard prosody such as sports commentary and anime: a two-stage pipeline in which an autoregressive transformer predicts coarse EnCodec features and a multinomial diffusion (DDPM) model fills in the rest, cloning a voice from a 2 to 12 second reference (about 6 seconds works best), with an optional 'deep clone' using the reference transcript and prosody steered by punctuation and capitalization. AGPL 3.0.
• Canva AI (Magic Studio) Canva · Design AI · 2023 — AI design tools. Image gen, editing, writing.
• Canva Magic Write Canva · AI Writing · 2023 — AI text generation in Canva.
• Captions AI Captions · Video Editing · 2023 — AI captions, eye contact, editing.
• Cerebras 1T Cerebras · LLM · 2025 — One-trillion-parameter model that Cerebras and Sandia National Laboratories trained on a single CS-3 system (with 55 TB of MemoryX memory) as a NeurIPS 2024 demonstration, then scaled to 16 CS-3s. It is not offered on Cerebras Inference.
• Cerebras Llama 3.3 70B Cerebras · LLM · 2024 — Llama 3.3 optimized for Cerebras hardware.
• Cerebras-GPT Cerebras · LLM · 2023 — Family of open, compute-optimal language models (111M to 13B parameters) trained on the Pile following Chinchilla scaling rules on Cerebras's wafer-scale cluster.
• Character.AI Character.AI · Chat Interface · 2022 — Conversational AI characters. Social platform.
• China Telecom AI T1 China Telecom AI · LLM · 2024 — Chinese telecom AI model.
• China Telecom Artificial Intelligence Research Institute T1 China Telecom AI · LLM · 2025 — Reasoning model in the TeleChat series from China Telecom's AI research institute (TeleAI), trained on domestic Chinese compute with curriculum-based fine-tuning and reinforcement learning. Open-sourced in April 2025 in two sizes, T1-35B and T1-115B.
• Civitai Civitai · Model Hub · 2023 — Community hub for SD models/LoRAs.
• Cleanvoice AI Cleanvoice · Audio Editing · 2023 — AI podcast audio cleanup.
• Cleo AI Cleo · Finance AI · 2023 — AI money management assistant.
• ClickUp Brain ClickUp · Productivity AI · 2023 — AI assistant across ClickUp workspace.
• CMU Bamba-9B CMU · LLM · 2025 — Mamba-2-based 9B decoder-only model trained on 2T tokens of Dolma v1.7 plus 200B curated tokens. Apache 2.0.
• Codeium Codeium · Code Completion · 2023 — Codeium, the AI coding company known for its free autocomplete extensions, renamed itself and its products Windsurf in April 2025, after its agentic IDE. In July 2025 Cognition, Devin's maker, agreed to buy Windsurf, then at $82M in annual recurring revenue with 350+ enterprise customers, and in June 2026 renamed the IDE Devin Desktop, a command center for local and cloud coding agents with a full IDE, unlimited Tab completions and Cognition's SWE-2 model.
• Devin Cognition AI · Code Agent · 2024 — Devin, Cognition's autonomous AI software engineer: it plans, writes, tests and ships production code in a team's codebase and tools, running in the cloud or locally, spinning up its own Windows, Mac and Linux virtual machines to build and test in a browser, and working until a pull request is ready to merge; teams can tag @Devin in Slack or Teams, and Devin Desktop (formerly Windsurf) manages local and cloud agents. Cognition says its annualized revenue passed $1B in September 2026.
• SWE-1.7 Cognition AI · Code LLM · 2026 — RL fine-tune of Moonshot's open Kimi K2.7 base (disclosed up front, unlike SWE-1.5's hidden GLM base), lifting FrontierCode 1.1 from 30.1% to 42.3% — tied with GPT-5.5 though behind Opus 4.8 — served at 1000 tokens per second.
• Aya 23 Cohere · LLM · 2024 — Cohere For AI's Aya 23, open-weight 8B and 35B multilingual models that pair a strong pre-trained model with the Aya instruction collection to serve 23 languages, covering about half the world's population; trading the 101-language breadth of Aya 101 for depth, Cohere reports they beat Aya 101 on the languages they cover as well as Gemma, Mistral and Mixtral on a wide range of tasks. Weights are for non-commercial use (CC BY-NC 4.0).
• Aya Expanse Cohere · LLM · 2024 — Cohere For AI's Aya Expanse, open-weight 8B and 32B multilingual models built from research on data arbitrage, multilingual preference training and model merging; on Arena-Hard-Auto translated into 23 languages they beat Gemma 2, Qwen 2.5 and Llama 3.1 in their size classes with up to a 76.6% win rate, and the 32B model beat the twice-as-large Llama 3.1 70B with a 54.0% win rate. Weights are for non-commercial use (CC BY-NC 4.0).
• Cohere Embed v3 Cohere · Embeddings · 2023 — Cohere's Embed v3 (November 2, 2023), embedding models for search and retrieval-augmented generation in English and multilingual (100+ languages, including cross-language search) versions with 1,024 dimensions plus light 384-dimension variants; trained to rank higher-quality documents first in noisy data and to hold up under vector compression, they led MTEB and BEIR zero-shot retrieval at launch, per Cohere, and Cohere's docs list a 512-token input limit and image input.
• Cohere Rerank 3 Cohere · Reranking · 2024 — Cohere's Rerank 3 (April 11, 2024), a reranking model for enterprise search and retrieval-augmented generation that reorders results from any search index or database by relevance, with a 4K-token context for long documents, ranking of semi-structured data such as emails, invoices, JSON, code and tables (including metadata like recency), coverage of 100+ languages and, per Cohere, up to 2x lower latency than Rerank 2 on short documents and 3x on long ones.
• Cohere Transcribe Arabic Cohere · Speech Recognition · 2026 — 2B-parameter Apache 2.0 Arabic speech-to-text: leads the Hugging Face Arabic ASR leaderboard at 25.87 WER (~11 points better than Whisper Large V3), with ~96% human preference head-to-head; handles dialect variety, code-switching, and Arabic-English bilingual speech.
• Command A Cohere · LLM · 2025 — Cohere's flagship enterprise model (111B parameters, 256K context) for tool use, RAG, agents and multilingual tasks; runs on two A100/H100 GPUs with 150% higher throughput than Command R+ 08-2024.
• Command R Cohere · LLM · 2024 — 35B. RAG-optimized. 128K context. Tool use. Multilingual.
• Command R+ Cohere · LLM · 2024 — 104B. Enterprise RAG. Multi-step tool use. On-premises deployment available.
• Command R7B Cohere · LLM · 2025 — Smallest model in Cohere's R series, built for speed and efficiency on commodity GPUs and edge devices (Dec 2024).
• Command-R (March 2024) Cohere · LLM · 2024 — Cohere's original Command R (March 11, 2024; c4ai-command-r-v01), a 35B-parameter generative model for production-scale enterprise use, optimized for retrieval-augmented generation and tool use with a 128K-token context, low latency and high throughput, strong capability in 10 key languages, and integration with Cohere's Embed and Rerank models; its weights were released on Hugging Face for research and evaluation under CC BY-NC 4.0.
• Colossyan Colossyan · Video Generation · 2022 — AI video with diverse avatars. Enterprise.
• OpenFold Columbia · Science/Biology · 2022 — OpenFold, a faithful but trainable open-source PyTorch reproduction of DeepMind's AlphaFold 2 protein structure prediction model, whose source code is Apache 2.0 licensed and which can also run DeepMind's pretrained AlphaFold parameters, now under CC BY 4.0.
• AceGPT Community · LLM · 2023 — AceGPT, open Arabic large language models from the FreedomIntelligence group, built on Llama 2 in 7B and 13B base and chat versions and aimed at Arabic's syntax and at alignment with local cultural values, through incremental pretraining on Arabic text, fine-tuning on genuine Arabic instructions with native GPT-4 responses and reinforcement learning from AI feedback (RLAIF); its authors report top results among open Arabic models on Arabic Vicuna-80, Arabic AlpacaEval, Arabic MMLU and EXAMs.
• Aider Community · Code Agent · 2023 — Aider, an open-source AI pair-programming tool that runs in the terminal and edits code in your local git repository: it connects to almost any LLM, cloud or local (its README cites Claude 3.7 Sonnet, DeepSeek R1 and V3, OpenAI o1, o3-mini and GPT-4o), builds a map of the whole codebase for context, auto-commits each change with a sensible message, works with 100+ programming languages, runs linters and tests after edits, and accepts images, web pages and voice requests.
• AnimateDiff Community · Video Generation · 2023 — AnimateDiff (ICLR 2024 Spotlight), a plug-and-play motion module that turns most community text-to-image models into animation generators without additional training; the main branch targets Stable Diffusion 1.5, with a beta branch for SDXL.
• AudioLDM 2 Community · Audio Generation · 2023 — AudioLDM 2 (August 2023), an open audio generation framework that uses one learning method for speech, music and sound effects: AudioMAE turns any audio into a 'language of audio' (LOA), a GPT-2 model translates text and other inputs into LOA, and a latent diffusion model conditioned on LOA generates the sound, with state-of-the-art or competitive results on text-to-audio, text-to-music and text-to-speech benchmarks. Its code also does super-resolution and inpainting; a 48 kHz model followed.
• Automatic1111 Community · Image Generation · 2022 — AUTOMATIC1111's Stable Diffusion web UI, a Gradio-based browser interface for Stable Diffusion with text-to-image and image-to-image modes, inpainting and outpainting, prompt attention weighting, textual-inversion training, X/Y/Z parameter plots and upscalers and face restorers such as GFPGAN, CodeFormer and RealESRGAN, plus a one-click install script.
• BioMistral Community · Medical LLM · 2024 — BioMistral, an open biomedical LLM built on Mistral 7B and further pretrained on PubMed Central, evaluated on 10 English medical question-answering tasks; its authors advise against using it in medical contexts without further evaluation.
• ChemBERTa Community · Science/Chemistry · 2020 — ChemBERTa (2020), a RoBERTa-style transformer from Seyone Chithrananda, Gabriel Grand and Bharath Ramsundar pre-trained as a masked language model on SMILES chemical strings and then fine-tuned to predict molecular properties; one of the first systematic tests of transformers on this task, it scaled well with pre-training data, gave competitive MoleculeNet results and attention maps of important substructures, and came with a 77M-SMILES PubChem pre-training set.
• ColPali Community · Retrieval · 2024 — ColPali, a visual document retriever that feeds PaliGemma-3B's image-patch outputs through a linear projection to build ColBERT-style multi-vector embeddings of document pages, so retrieval works directly on page images, covering text plus layout and charts, without OCR or layout-recognition pipelines; the repository also has later ColVision and bi-encoder retrievers.
• ComfyUI Community · Image Generation · 2023 — ComfyUI, an open-source (GPL-3.0) node-graph engine for generative AI that builds workflows for images, video, 3D and audio from the latest open models, plus partner nodes for closed models such as Nano Banana and Seedance; it focuses on local consumer hardware, can run large models on as little as 4 GB of VRAM and 8 GB of RAM through weight streaming, supports NVIDIA, AMD, Intel, Apple Silicon and Ascend GPUs, and ships as a desktop app, portable install or Comfy Cloud, with API endpoints.
• Continue Community · Code AI · 2023 — Continue, an open-source (Apache 2.0) coding agent from Continue Dev, Inc. for VS Code, JetBrains IDEs and the command line. Its team shipped a final 2.0.0 release of all three, removing anonymous telemetry and authentication, then made the GitHub repository read-only; Continue's website says the company was acquired by Cursor and that its open-source codebase remains freely available as a foundation for others.
• ControlNet Community · Image Control · 2023 — ControlNet, a neural-network structure that adds extra conditions to text-to-image diffusion models: it trains a copy of the model's blocks while the original weights stay locked, so even small datasets don't damage the base model; the released Stable Diffusion 1.5 versions condition on edges, lines, scribbles, human pose, segmentation, depth, normal maps and anime line art.
• CroissantLLM Community · LLM · 2024 — CroissantLLM, a 1.3B-parameter bilingual French-English model pretrained on 3T tokens with a 1:1 English-to-French ratio and a custom tokenizer, built to run fast on consumer hardware; its authors released the training data, code, dozens of checkpoints, chat and translation models and a new French benchmark, FrenchBench.
• Diffusion Bee Community · Image Generation · 2022 — DiffusionBee, a free, open-source (AGPL-3.0) macOS app that runs Stable Diffusion locally and offline with a one-click installer: text to image, image to image, generative fill (inpainting and outpainting), upscaling, image variants, control images, video tools and local training of custom models, with SD 1.x, 2.x and XL, ControlNet and LoRA support. It is optimized for Apple Silicon and also runs on Intel Macs, and prompts and images never leave the device.
• Dolphin 2.5 Mixtral Community · LLM · 2024 — Eric Hartford's Dolphin 2.5, a fine-tune of Mixtral-8x7b trained with a 16K context on a dataset mix heavy in code (adding Synthia, OpenHermes, PureDove and a new Dolphin-Coder set); its card says the dataset was filtered to remove alignment, making the model uncensored and highly compliant, and advises adding your own alignment layer before serving it.
• Draw Things Community · Image Generation · 2023 — Draw Things, from Draw Things, Inc., an app for iPhone, iPad and Mac that generates images locally and offline with downloadable models, plus image-to-video, an infinite canvas for expanding and filling images, restyling of photos, virtual try-on and on-device LoRA training. The core app is free; a free account adds cloud compute, and Draw Things+ ($8.99 a month) raises cloud limits and allows custom LoRAs on cloud compute.
• F5-TTS Community · Speech Synthesis · 2024 — F5-TTS, an open text-to-speech system built as a flow-matching diffusion transformer with ConvNeXt V2 for faster training and inference, plus Sway Sampling, an inference-time strategy that improves quality; it speaks in the voice of a supplied reference clip and its transcript. Code is MIT, while the pretrained models are CC BY-NC because of their Emilia training data.
• FinGPT Community · Finance LLM · 2023 — The AI4Finance Foundation's FinGPT, an open-source project for financial large language models that adapts existing open LLMs with lightweight fine-tuning (such as LoRA) instead of training from scratch, so models can take in new financial data for under $300 per fine-tuning, with an automatic data-curation pipeline for internet-scale financial data; releases include LoRA-tuned financial sentiment models and FinGPT-Forecaster (November 2023).
• Flowise Community · AI Platform · 2023 — Flowise, FlowiseAI's open-source (Apache 2.0) platform for building AI agents visually: Agentflow for multi-agent systems, Chatflow for chatbots with tool calling and RAG, human-in-the-loop review, tracing, APIs, SDKs and embeddable chat, with 100+ LLMs, embeddings and vector databases. Citing developers' shift to coding agents, FlowiseAI froze development on July 29, 2026, archived the GitHub repository on August 10 and ended official support on August 31, 2026.
• Fooocus Community · Image Generation · 2023 — Fooocus, an offline, open-source (GPL-3.0) image generator by lllyasviel, built on Stable Diffusion XL and Gradio and designed like Midjourney so users focus on prompts, not settings: an offline GPT-2-based prompt engine, its own inpainting and image-prompt algorithms, upscaling, variations, styles and Civitai SDXL models, a sub-three-click install and a 4 GB Nvidia VRAM minimum. It is now in limited long-term support with bug fixes only and no plans to adopt newer architectures such as Flux.
• Gaussian Splatting Community · 3D Rendering · 2023 — 3D Gaussian Splatting (SIGGRAPH 2023), a radiance-field method from Inria, Université Côte d'Azur and the Max Planck Institute for Informatics (Bernhard Kerbl, Georgios Kopanas, Thomas Leimkühler and George Drettakis) that represents a photographed scene as optimized 3D Gaussians and draws them with a fast visibility-aware rasterizer, giving state-of-the-art quality with competitive training times and, unlike earlier methods, real-time (30+ fps) novel-view rendering of complete scenes at 1080p.
• GeoGPT Community · Science/Earth · 2023 — GeoGPT, a geoscience foundation model and research platform (run from Zhejiang Lab in Hangzhou) that combines large language models with AI tools: it reads and reasons over high-resolution geological maps, offers Vibe Research and Agent Builder tools for scientists on its zero2x platform, and has supported projects such as a deep-time timeline built from 20,000+ geological profiles. Open weights are available for Llama3.1-70B-GeoGPT, Qwen2.5-72B-GeoGPT and GeoGPT-R1-Preview.
• InstantID Community · Image Generation · 2024 — InstantID, a tuning-free method from the InstantX team and Xiaohongshu that generates identity-preserving images from a single face photo in seconds and supports various downstream tasks; its pipeline was merged into diffusers and works with LCM acceleration and multiple ControlNets.
• IP-Adapter Community · Image Control · 2023 — Tencent AI Lab's IP-Adapter, a 22M-parameter add-on that lets pretrained text-to-image diffusion models such as Stable Diffusion 1.5 and SDXL take an image as the prompt, using decoupled cross-attention that keeps image and text features separate so the two kinds of prompt can be combined; it matches or beats fully fine-tuned image-prompt models, carries over to custom models built on the same base and to ControlNet-style controls, and later gained FaceID variants. Apache 2.0 weights.
• Kokoro TTS Community · Speech Synthesis · 2024 — Kokoro, an open-weight 82M-parameter text-to-speech model built on the StyleTTS 2 and ISTFTNet architectures (decoder only, no diffusion), with Apache 2.0 weights; its developer says it matches larger models' quality while being much faster and cheaper, and v1.0 (Jan 27, 2025) covers 8 languages and 54 voices.
• LangFlow Community · AI Platform · 2023 — Langflow, an open-source (MIT) low-code platform for building and deploying AI agents, workflows and MCP servers: a visual drag-and-drop builder whose components can be customized in Python, an interactive playground, multi-agent orchestration with retrieval, deployment as an API, JSON export or MCP server, observability through LangSmith and Langfuse, and support for all major LLMs and vector databases. It runs as a Python package, as Langflow Desktop for Windows and macOS, or in the cloud.
• llama.cpp Community · Inference Engine · 2023 — llama.cpp, a dependency-free C/C++ engine built on the ggml library for running LLMs and vision-language models locally or in the cloud, with 1.5-bit to 8-bit integer quantization, Apple silicon (Metal), x86 and RISC-V CPU optimizations, CUDA, HIP, Vulkan, SYCL and other GPU backends, hybrid CPU+GPU inference for models larger than VRAM, a CLI and an OpenAI-compatible server.
• MAP-Neo Community · LLM · 2024 — MAP-Neo, a fully open-source 7B language model from the Multimodal Art Projection (M-A-P) community, released with its pretraining data, the Matrix data-processing pipeline, pretraining scripts, alignment code and intermediate checkpoints; trained from scratch on 4.5 trillion English and Chinese tokens, it performs comparably to Llama 2 7B and does well on reasoning, math and coding for its size. Commercial use is permitted.
• ModernBERT Community · LLM · 2024 — ModernBERT (December 2024) from Answer.AI and LightOn, a modernized BERT-style encoder-only model pre-trained on 2 trillion tokens of English and code with a native 8,192-token context, using rotary position embeddings, alternating local-global attention, unpadding and Flash Attention; in base (149M parameters) and large (395M) sizes, its authors report state-of-the-art classification and retrieval results, including on code, and the best speed and memory efficiency of any encoder. Apache 2.0.
• Open WebUI Community · Chat Interface · 2024 — Open WebUI, a self-hosted, extensible AI platform built to run entirely offline that connects to Ollama and OpenAI-compatible APIs such as LM Studio, Groq, Mistral, OpenRouter and vLLM: it adds role-based access control, plugins (filters, actions, pipes, tools and skills), MCP and OpenAPI tool servers, custom agents, an agent terminal, notes, team channels, memory, calendars and automations, installs with pip, Docker or Kubernetes, and reports 417M+ downloads. An enterprise plan is offered.
• Open-Sora Community · Video Generation · 2024 — HPC-AI Tech's Open-Sora, an open-source initiative for efficient, high-quality video generation that releases its models, tools and training details; Open-Sora 2.0 (11B parameters, March 2025) matches the 11B HunyuanVideo and 30B Step-Video on VBench and human preference, with checkpoints and training code for training at about $200K, following the 1B Open-Sora 1.3 (February 2025).
• OpenChat 3.5 Community · LLM · 2023 — OpenChat 3.5, a 7B open chat model with an 8,192-token context, fine-tuned with C-RLFT, a method inspired by offline reinforcement learning that learns from mixed-quality data without preference labels, on public instruction datasets such as ShareGPT, OpenOrca and MetaMathQA; its developers reported a 7.81 MT-Bench score, which they called the best among open-source models at release, ahead of 70B models.
• OuteTTS Community · Speech Synthesis · 2024 — OuteAI's OuteTTS, an open text-to-speech model family; version 1.0 1B (Llama-OuteTTS-1.0-1B) continues training Llama 3.2 1B on about 60K hours of audio with a DAC audio encoder, clones a voice from about 10 seconds of reference audio, aligns raw text automatically and speaks many languages, strongest in English, Arabic, Chinese, Dutch, French, German, Italian, Japanese, Korean, Lithuanian, Russian and Spanish. Non-commercial license (CC BY-NC-SA 4.0 plus Llama 3.2 terms).
• PhotoMaker Community · Image Editing · 2023 — Tencent ARC's PhotoMaker, which customizes realistic human photos from reference images through a stacked ID embedding, within seconds and without LoRA training, aiming for high identity fidelity with diverse, text-controllable output; it can act as an adapter with other base models and LoRA modules, and PhotoMaker V2 followed with support from the HunyuanDiT team.
• Piper TTS Community · Speech Synthesis · 2023 — Piper, a fast, local neural text-to-speech engine that uses espeak-ng for phonemization, with a command-line tool, web server, Python and C/C++ APIs and support for training new voices; its voice list covers 39 languages, and it is used by Home Assistant, the NVDA screen reader and other projects. Development moved from rhasspy/piper to the Open Home Foundation's GPL-3.0 piper1-gpl repository, installable with 'pip install piper-tts'.
• Saiga Community · LLM · 2023 — Saiga, Ilya Gusev's Russian-language chatbot models fine-tuned from open base models; Saiga/Llama3 8B, for example, is built on Llama 3 8B Instruct and trained with supervised fine-tuning and KTO preference optimization, using training scripts published in the Saiga GitHub repository.
• StyleTTS 2 Community · Speech Synthesis · 2023 — StyleTTS 2 (June 2023), an open text-to-speech model by Yinghao Aaron Li, Nima Mesgarani et al. that models speaking style as a latent variable sampled by diffusion, needing no reference audio, and trains end to end with large speech language models such as WavLM as discriminators; native English listeners rated it above human recordings on LJSpeech and on par on VCTK, and it beat earlier public models at zero-shot speaker adaptation. Code MIT; weights require disclosing synthetic speech.
• TinyLlama Community · LLM · 2024 — TinyLlama, a compact 1.1B-parameter model pretrained on 3 trillion tokens (training began Sept 1, 2023, on 16 A100-40G GPUs) with exactly the same architecture and tokenizer as Llama 2, so it drops into Llama-based projects; aimed at uses with tight compute and memory budgets.
• Tortoise TTS Community · Speech Synthesis · 2022 — Tortoise (TorToiSe), a text-to-speech program built for strong multi-voice capabilities and highly realistic prosody and intonation, combining an autoregressive decoder with a diffusion decoder; its author named it for being very slow (originally about one medium-sized sentence every 2 minutes on a K80), but it now reaches a 0.25-0.3 real-time factor on 4GB of VRAM and under 500 ms latency with streaming.
• Wonder3D Community · 3D Generation · 2023 — Wonder3D (CVPR 2024 Highlight), which reconstructs highly detailed textured meshes from a single image in 2 to 3 minutes: a cross-domain diffusion model generates consistent multi-view normal maps and color images, then a normal fusion method reconstructs the mesh; its authors have since released Wonder3D++.
• Zeroscope v2 Community · Video Generation · 2023 — Watermark-free ModelScope-based text-to-video model (zeroscope_v2 576w), fine-tuned on 9,923 clips and 29,769 tagged frames at 24 frames and 576x320, meant for drafting before upscaling with zeroscope_v2_XL; it uses about 7.9 GB of VRAM for 30 frames.
• Consensus AI Consensus · Research AI · 2023 — AI-powered academic paper search.
• Consortium Spectra 1.1 Consortium · LLM · 2025 — Ternary language models (TriLMs) trained with quantization-aware training, from a study of their scaling laws and efficient inference; the authors find TriLMs gain more from extra training data than from extra parameters.
• Copy.ai Copy.ai · AI Writing · 2023 — AI copywriting and marketing content.
• Coqui TTS Coqui · Speech Synthesis · 2023 — Coqui's TTS, an open-source library for advanced text-to-speech with pretrained models in over 1,100 languages, tools for training and fine-tuning models in any language and utilities for dataset analysis; it includes the XTTS v2 voice-cloning model, with fine-tuning code and streaming.
• XTTS v2 Coqui · Speech Synthesis · 2023 — Coqui's XTTS v2, a voice generation model that clones a voice into other languages from a 6-second audio clip, supporting 17 languages, emotion and style transfer, cross-language cloning, multiple speaker references and 24 kHz output; compared with XTTS v1 it adds Hungarian and Korean and improves speaker conditioning, stability, prosody and audio quality.
• DenseNet Cornell/Tsinghua · Vision/CV · 2017 — DenseNet (Densely Connected Convolutional Networks), which connects each layer to every later layer, so a network with L layers has L(L+1)/2 direct connections; this eases vanishing gradients, strengthens feature propagation and reuse and cuts parameters, and DenseNets improved on the state of the art on most of CIFAR-10, CIFAR-100, SVHN and ImageNet with less computation. Code and pretrained models are on GitHub.
• Craiyon (DALL-E Mini) Craiyon · Image Generation · 2022 — Free text-to-image. Originally DALL-E Mini.
• CrowdStrike Charlotte AI CrowdStrike · Security AI · 2023 — AI security analyst.
• D-ID D-ID · Video Generation · 2022 — Talking avatar creation from photos.
• DBRX Instruct Databricks · LLM (MoE) · 2024 — 132B MoE, 36B active. Open-source. Surpassed Mixtral and Llama 2 70B at launch.
• Dolly 2.0 Databricks · LLM · 2023 — Databricks' Dolly 2.0 (April 12, 2023), a 12B-parameter instruction-following model fine-tuned from EleutherAI's Pythia-12B on databricks-dolly-15k, 15,000 prompt-and-response pairs written by more than 5,000 Databricks employees; Databricks called it the first open-source instruction-following LLM tuned on human-generated data licensed for research and commercial use, and released its weights, training code and dataset (CC BY-SA 3.0), while noting it is not a state-of-the-art model.
• DBRX Databricks and Mosaic ML · LLM · 2024 — Databricks MoE model. 132B total, 36B active. Open-source. Enterprise-optimized.
• DataRobot AI DataRobot · ML Platform · 2023 — Enterprise AI/ML platform.
• IF (DeepFloyd) DeepFloyd · Image Generation · 2023 — DeepFloyd IF from DeepFloyd Lab at Stability AI, an open-source text-to-image model built from a frozen T5 text encoder and three cascaded pixel diffusion modules: a base model that generates 64x64 images and two super-resolution modules up to 256x256 and 1024x1024; its developers report a zero-shot COCO FID of 6.66, and the 4.3B IF-I-XL base plus 1.2B IF-II-L upscaler need 16GB of VRAM.
• DeepL DeepL · Translation · 2017 — Neural machine translation. 31 languages.
• DeepL Write DeepL · AI Writing · 2023 — AI writing assistant for style improvement.
• AlphaFold 2 DeepMind · Science/Biology · 2020 — AlphaFold 2, the DeepMind AI system that predicts a protein's 3D structure from its amino-acid sequence, which CASP's organizers recognized in November 2020 as a solution to biology's 50-year-old protein folding problem: at CASP14 it reached a median 92.4 GDT, with errors of about 1.6 angstroms. DeepMind open-sourced it in 2021, its database with EMBL-EBI grew to over 200 million predicted structures, and Demis Hassabis and John Jumper shared the 2024 Nobel Prize in Chemistry for it.
• AlphaGo DeepMind · Game AI · 2016 — Google DeepMind's AlphaGo, a Go-playing system that combines deep neural networks, a policy network that picks moves and a value network that predicts the winner, with advanced search, trained first on thousands of expert human games and then by reinforcement learning against versions of itself; it beat European champion Fan Hui 5-0 in October 2015, the first match won by an AI against a Go professional, and Lee Sedol 4-1 in Seoul in March 2016, earning a 9 dan professional ranking.
• AlphaZero DeepMind · Game AI · 2017 — DeepMind's AlphaZero, introduced in late 2017 and fully evaluated in Science in December 2018, a single system that taught itself chess, shogi and Go from random play through self-play reinforcement learning, given only the rules, with a deep neural network guiding Monte Carlo tree search; after about 9 hours of training for chess, 12 for shogi and 13 days for Go, it beat the leading programs Stockfish (155 wins and 6 losses in 1,000 games), Elmo (91.2% of games) and AlphaGo Zero (61%).
• Chinchilla DeepMind · LLM · 2022 — Chinchilla (March 2022), DeepMind's 70-billion-parameter language model, built to test its finding from training over 400 models that model size and training data should grow in equal proportion and that large models of the day were undertrained. Trained with the same compute as the 280B Gopher but a quarter of its size and four times the data, Chinchilla outperformed Gopher, GPT-3, Jurassic-1 and Megatron-Turing NLG on most tasks, reached 67.5% on MMLU and cost less to fine-tune and run.
• Gato DeepMind · RL/Agents · 2022 — Google DeepMind's Gato (May 2022), a 1.2-billion-parameter generalist agent: a single Transformer with one set of weights, trained on 604 distinct tasks, that can play Atari, caption images, chat and stack blocks with a real robot arm, deciding from context whether to output text, joint torques, button presses or other tokens. It performed over 450 of the 604 tasks at above 50% of expert score.
• Gopher DeepMind · LLM · 2021 — Google DeepMind's Gopher (December 8, 2021), a 280-billion-parameter Transformer language model, the largest in a series starting at 44 million parameters, evaluated on 152 tasks with state-of-the-art results on most; scale helped most in reading comprehension, fact-checking and identifying toxic language and least in logical and mathematical reasoning, it moved toward human expert performance on MMLU, and DeepMind noted failure modes such as repetition, bias and confident errors.
• MuZero DeepMind · RL/Agents · 2019 — Google DeepMind's MuZero, a successor to AlphaGo and AlphaZero that masters Go, chess, shogi and Atari without being told the rules: it learns a model of only what matters for planning (value, policy and reward) and combines it with AlphaZero's lookahead tree search, setting a new state of the art on 57 Atari games while matching AlphaZero's superhuman play in Go, chess and shogi. First described in a 2019 preprint, it was published in Nature in December 2020.
• DeepSeek Chat DeepSeek · Chat Interface · 2024 — DeepSeek's free AI assistant on the web and in mobile apps (the apps launched January 15, 2025, with no ads or in-app purchases), offering DeepThink reasoning, web search, file upload with text extraction and synced chat history; DeepSeek-V4.1-Flash-1017 went live on the web and app on October 17, 2026.
• DeepSeek Coder V2 DeepSeek · Code LLM · 2024 — DeepSeek's DeepSeek-Coder-V2, an open mixture-of-experts code model with 236B total and 21B active parameters (plus a 16B Lite version with 2.4B active), further pretrained from a DeepSeek-V2 checkpoint on 6 trillion more tokens, which expanded support from 86 to 338 programming languages and context from 16K to 128K; DeepSeek reports it beats GPT-4 Turbo, Claude 3 Opus and Gemini 1.5 Pro on coding and math benchmarks. Commercial use is allowed under the DeepSeek Model License.
• DeepSeek-Coder DeepSeek · Code LLM · 2023 — DeepSeek's DeepSeek Coder, a family of open code models from 1B to 33B parameters, each trained from scratch on 2 trillion tokens (87% code, 13% English and Chinese natural language) with a 16K window and a fill-in-the-blank task for project-level code completion and infilling; DeepSeek reported state-of-the-art results among open code models on HumanEval, MultiPL-E, MBPP, DS-1000 and APPS, and the models allow commercial use under the DeepSeek Model License.
• DeepSeek-Coder-V2-Lite DeepSeek · Code LLM · 2024 — DeepSeek-Coder-V2-Lite, the 16B-parameter version (2.4B active parameters, 128K context) of DeepSeek's open mixture-of-experts code model DeepSeek-Coder-V2, which was further pretrained from a DeepSeek-V2 checkpoint on 6 trillion more tokens and supports 338 programming languages; DeepSeek reports DeepSeek-Coder-V2 beats GPT-4 Turbo, Claude 3 Opus and Gemini 1.5 Pro on coding and math benchmarks.
• DeepSeek-GRM-27B DeepSeek · LLM · 2025 — 27B generalist reward model from DeepSeek's research on inference-time scaling for reward modeling: its reward judgments on general queries improve as more inference compute is spent.
• DeepSeek-LLM DeepSeek · LLM · 2024 — DeepSeek LLM, DeepSeek's open-weight general-purpose language models: 7B and 67B base and chat versions trained from scratch on 2 trillion English and Chinese tokens, with the chat models tuned by supervised fine-tuning and DPO, licensed for commercial use. DeepSeek reports that the 67B base beats Llama 2 70B on reasoning, coding, math and Chinese, and that 67B Chat scores 73.78 on HumanEval and 84.1 on GSM8K and outperforms GPT-3.5 in Chinese and open-ended evaluations.
• DeepSeek-R1 DeepSeek · Reasoning LLM · 2025 — DeepSeek-R1, DeepSeek's first-generation open reasoning model (MIT license), built on the 671B-parameter DeepSeek-V3-Base: where its sibling DeepSeek-R1-Zero learned to reason through reinforcement learning alone, R1 adds cold-start data before large-scale RL to fix repetition, readability and language mixing, and DeepSeek says it matches OpenAI o1 on math, code and reasoning. It came with six distilled models (1.5B to 70B) and launched on the API at $0.55/$2.19 per million tokens.
• DeepSeek-R1-0528 DeepSeek · Reasoning LLM · 2025 — DeepSeek's DeepSeek-R1-0528, a May 28, 2025 upgrade of its open-weight DeepSeek-R1 reasoning model: more compute and algorithmic optimizations in post-training deepened its reasoning (about 23K tokens per AIME question, up from 12K), lifting AIME 2025 accuracy from 70% to 87.5% and bringing it close to o3 and Gemini 2.5 Pro; it also cut hallucinations, improved front-end coding, added system-prompt support, JSON output and function calling, and was distilled into DeepSeek-R1-0528-Qwen3-8B.
• DeepSeek-R1-Distill-Llama-8B DeepSeek · Reasoning LLM · 2025 — Llama 3.1 8B base model fine-tuned on reasoning samples generated by DeepSeek-R1; DeepSeek reports 50.4% on AIME 2024 (pass@1) and 89.1% on MATH-500.
• DeepSeek-R1-Distill-Qwen-7B DeepSeek · Reasoning LLM · 2025 — Qwen2.5-Math-7B fine-tuned on reasoning samples generated by DeepSeek-R1; DeepSeek reports 55.5% on AIME 2024 (pass@1) and 92.8% on MATH-500.
• DeepSeek-V2 DeepSeek · LLM · 2024 — DeepSeek's DeepSeek-V2, a mixture-of-experts language model with 236B total parameters (21B active per token) and a 128K context, pretrained on 8.1 trillion tokens and tuned with SFT and RL; DeepSeek reports it beats its dense DeepSeek 67B while cutting training costs by 42.5%, shrinking the KV cache by 93.3% and raising maximum generation throughput 5.76-fold, and it offered the model through an OpenAI-compatible API at what it called an unbeatable price. Commercial use is allowed.
• DeepSeek-V2.5 DeepSeek · LLM (MoE) · 2024 — DeepSeek's DeepSeek-V2.5 (September 2024), which merged DeepSeek-V2-Chat-0628 and DeepSeek-Coder-V2-0724 into one model with both general chat and coding ability, better alignment with human preferences and improved writing and instruction following; built on DeepSeek-V2's mixture-of-experts design, it scored 50.5 on AlpacaEval 2.0, 76.2 on ArenaHard and 89 on HumanEval Python, and served both the deepseek-chat and deepseek-coder API names.
• DeepSeek-V3 DeepSeek · LLM · 2024 — DeepSeek-V3 (December 2024), DeepSeek's open-weight mixture-of-experts model with 671B total and 37B active parameters (685B with its multi-token prediction module), a 128K-token context and 14.8T training tokens; it uses Multi-head Latent Attention, auxiliary-loss-free load balancing and multi-token prediction, and its full training took only 2.788M H800 GPU hours. DeepSeek reports it beats other open models and rivals leading closed ones while generating 60 tokens a second, 3x faster than V2.
• DeepSeek-V3 0324 DeepSeek · LLM · 2025 — DeepSeek's DeepSeek-V3-0324, a March 2025 update of DeepSeek-V3 with much stronger reasoning (MMLU-Pro 75.9 to 81.2, GPQA 59.1 to 68.4, AIME 39.6 to 59.4, LiveCodeBench 39.2 to 49.2), more executable and better-looking front-end web code, better Chinese writing and search-report analysis, and more accurate function calling; DeepSeek released its weights under the MIT license, as with DeepSeek-R1.
• DeepSeek-V3.1 DeepSeek · LLM · 2025 — Hybrid model with switchable thinking and non-thinking modes; stronger tool use and agent tasks, with thinking-mode answers comparable to DeepSeek-R1-0528 but faster. Open weights (MIT).
• DeepSeek-V3.1-Terminus DeepSeek · LLM · 2025 — Sep 2025 update to DeepSeek-V3.1 that cuts Chinese/English mix-ups and stray characters and strengthens its Code Agent and Search Agent. Open weights.
• DeepSeek-V3.2 DeepSeek · Reasoning LLM · 2025 — DeepSeek's reasoning-first model, released Dec 1, 2025 as the official successor to V3.2-Exp alongside the higher-reasoning V3.2-Speciale; its first model to build thinking directly into tool use. Open weights under MIT; DeepSeek's API has since moved to the V4 family.
• DeepSeek-V3.2-Exp DeepSeek · LLM · 2025 — Experimental step toward DeepSeek's next architecture: built on V3.1-Terminus with DeepSeek Sparse Attention (DSA) for cheaper long-context training and inference; benchmarks on par with V3.1-Terminus (MMLU-Pro 85.0, AIME 2025 89.3, SWE-bench Verified 67.8).
• DeepSeek-V3.2-Speciale DeepSeek · Reasoning LLM · 2025 — Maxed-out reasoning variant released with DeepSeek-V3.2 on Dec 1, 2025. DeepSeek reports gold-level results at IMO, CMO, ICPC World Finals and IOI 2025 and says it rivals Gemini 3.0 Pro, at the cost of higher token usage. Fine-tuned from DeepSeek-V3.2-Exp-Base.
• DeepSeek-V4-Flash DeepSeek · LLM · 2026 — Compact V4 variant (160GB weights), 1M-token context, released alongside V4 Pro under MIT license. Defines a new ultra-low-cost tier - output tokens ~1/89th the cost of Claude Opus 4.7. Official release Jul 31, 2026 (V4-Flash-0731): 284B/13B-active, Terminal-Bench 2.1 82.7 (above V4-Pro-Preview), 1M context, native Responses-API protocol; MIT-licensed weights on Hugging Face.
• DeepSeek-V4-Flash-Vision-Exp DeepSeek · Multimodal LLM · 2026 — DeepSeek-V4-Flash-Vision-Exp (August 21, 2026), an experimental multimodal version of DeepSeek-V4-Flash that DeepSeek says matched V4-Flash on text, agents and reasoning while bringing multimodal agent performance close to Claude Opus 4.8; images counted as up to 384 tokens each at V4-Flash prices, and its weights were published on Hugging Face. It was retired on September 10, 2026, when DeepSeek-V4.1-Flash took over its API name.
• DeepSeek-V4-Pro DeepSeek · LLM · 2026 — Largest open-weights model at release (~1.6T total params, 49B active; 865GB on Hugging Face), 1M-token context. SWE-bench Verified 80.6% (vendor-reported), competitive with frontier proprietary models at a fraction of the price. Released alongside V4 Flash.
• DeepSeek-V4.1-Flash DeepSeek · Multimodal LLM · 2026 — DeepSeek-V4.1-Flash (September 10, 2026), DeepSeek's open-weight (MIT) multimodal mixture-of-experts model with 552B backbone parameters and a 1M-token context: its causal encoder-decoder activates 8B parameters per token in prefill and 16B in decode, and sparse attention with FP4 KV caching cuts the KV cache to about a quarter of V4-Flash's. The V4.1-Flash-1017 update (October 17, 2026), served as deepseek-flash, brought stronger benchmarks, faster task completion and shorter outputs.
• DeepSeek-VL DeepSeek · Multimodal LLM · 2024 — DeepSeek's DeepSeek-VL, an open vision-language model for real-world understanding of logical diagrams, web pages, formulas, scientific literature, natural images and embodied scenes; the 7B version pairs a hybrid SigLIP-L and SAM-B vision encoder (1024x1024 input) with DeepSeek-LLM-7B and was trained on about 400B vision-language tokens, with an instruction-tuned chat version. Commercial use is allowed under the DeepSeek Model License.
• DeepSeek-VL2 DeepSeek · Multimodal LLM · 2024 — DeepSeek's DeepSeek-VL2, a series of mixture-of-experts vision-language models in Tiny, Small and standard versions with 1.0B, 2.8B and 4.5B activated parameters (the standard model built on DeepSeekMoE-27B); DeepSeek reports competitive or state-of-the-art results with similar or fewer activated parameters than other open models on visual question answering, OCR, document, table and chart understanding and visual grounding.
• Descript Descript · Audio/Video Editing · 2022 — Text-based audio/video editing. Overdub voice.
• Dify Dify · AI Platform · 2023 — Open-source LLM app development platform.
• Discord Clyde Discord · Chat Interface · 2023 — Experimental AI chatbot built into Discord using OpenAI models; Discord shut it down on December 1, 2023.
• DoNotPay DoNotPay · Legal AI · 2023 — AI consumer law assistant.
• Duolingo Max Duolingo · Education AI · 2023 — GPT-4 powered language tutoring.
• Elai.io Elai · Video Generation · 2022 — AI video from text. 75+ avatars.
• Comma v0.1-2T EleutherAI · LLM · 2025 — 7B base model trained on 2 trillion tokens of openly licensed text from the Common Pile; it performs comparably to budget-matched models trained on unlicensed data. Apache 2.0.
• GPT-J EleutherAI · LLM · 2021 — EleutherAI's GPT-J 6B, a 6-billion-parameter English language model (28 layers, 2,048-token context, rotary position embeddings) trained with Ben Wang's Mesh Transformer JAX on 402 billion tokens of the Pile on a TPU v3-256 pod; EleutherAI says it is not meant to be deployed without fine-tuning, supervision or moderation.
• GPT-J-6B EleutherAI · LLM · 2021 — EleutherAI's GPT-J 6B, a 6-billion-parameter English language model (28 layers, 2,048-token context, rotary position embeddings) trained with Ben Wang's Mesh Transformer JAX on 402 billion tokens of the Pile on a TPU v3-256 pod; EleutherAI says it is not meant to be deployed without fine-tuning, supervision or moderation.
• GPT-Neo EleutherAI · LLM · 2021 — EleutherAI open-source GPT alternative. Community-built. 2.7B params.
• GPT-NeoX EleutherAI · LLM · 2022 — EleutherAI's 20B-parameter autoregressive language model (GPT-NeoX-20B), trained on the Pile; at release, the largest dense model with publicly available weights. Open weights (Apache 2.0).
• Pythia EleutherAI · LLM · 2023 — EleutherAI's Pythia, a suite of language models for research on interpretability, learning dynamics and transparency: eight sizes from 70M to 12B parameters (later plus 14M and 31M), each trained on the Pile or a deduplicated Pile in the same order for about 300B tokens, with 154 checkpoints per model and all models, data and code released.
• Eleven Music ElevenLabs · Music Generation · 2026 — ElevenLabs' entry into AI music generation: studio-quality vocal and instrumental tracks from text prompts, integrated with its voice/audio stack. Competes directly with Suno and Udio.
• ElevenLabs ElevenLabs · Voice / TTS · 2023 — Leading AI voice platform. Voice cloning. Text-to-speech. Sound effects. Near-human quality.
• ElevenLabs Turbo v2 ElevenLabs · Speech Synthesis · 2024 — First-generation low-latency text-to-speech model (English only); ElevenLabs now lists it as deprecated and recommends Eleven Flash v2 instead.
• ElevenLabs v3 ElevenLabs · Speech Synthesis · 2025 — Eleven v3 (alpha, June 2025): expressive text-to-speech with audio tags for emotion and delivery, multi-speaker dialogue through the Text to Dialogue API, and 70+ languages. ElevenLabs now calls it its previous-generation model; Eleven v4 is its flagship.
• Elicit Elicit · Research AI · 2023 — AI research assistant. Paper analysis.
• ELYZA-japanese-Llama-2-7b ELYZA · LLM · 2023 — ELYZA's ELYZA-japanese-Llama-2-7b, Meta's Llama 2 7B with additional pretraining to extend its Japanese ability (6.27B parameters), released with an instruction-tuned version and 'fast' versions with an expanded 45,043-token vocabulary; under the Llama 2 Community License.
• Meditron EPFL · Medical LLM · 2023 — EPFL's MEDITRON, open-source 7B and 70B medical LLMs built on Llama 2 with continued pretraining on a curated medical corpus of selected PubMed articles, abstracts and international medical guidelines; EPFL reports a 6% absolute gain over the best public baseline in its class, and MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and comes within 5% of GPT-4 and 10% of Med-PaLM 2.
• SaulLM-7B Equall AI · Legal LLM · 2024 — Equall's SaulLM-7B (Saul-7B-Instruct), an English legal-domain instruction model built by continued pretraining of Mistral-7B, developed with CentraleSupélec, Sorbonne Université, Instituto Superior Técnico and NOVA School of Law.
• Apertus ETH Zurich and EPF Lausanne · LLM · 2025 — Fully open multilingual LLM from the Swiss AI Initiative in 8B and 70B sizes, supporting over 1,000 languages and trained on 15T tokens of compliant open data. Apache 2.0.
• Exa AI Exa · Search AI · 2023 — Neural search API for AI applications.
• ResNeXt-101 Facebook · Vision/CV · 2017 — Facebook AI Research's ResNeXt-101, a 101-layer ResNeXt network; ResNeXt builds networks by repeating a block that aggregates a set of same-topology transformations, adding 'cardinality' (the number of transformations) as a dimension beside depth and width. Increasing cardinality improved ImageNet accuracy more than going deeper or wider, and ResNeXt was the basis of an entry that placed second in ILSVRC 2016 classification.
• wave2vec 2.0 LARGE Facebook · LLM · 2020 — Self-supervised speech model that learns from raw audio (masking speech in latent space and solving a contrastive task) before fine-tuning on transcripts; it set LibriSpeech records while needing far less labeled data.
• DINOv2 Facebook AI Research · Vision/CV · 2023 — DINOv2. Facebook AI Research Vision/CV model.
• ULMFiT fast.ai · Transfer Learning · 2018 — Universal Language Model Fine-tuning. Transfer learning pioneer.
• Featherless AI Qwerky-72B Featherless AI · LLM · 2025 — Featherless AI's conversion of Qwen2.5-72B-Instruct to RWKV-style linear attention, which drops quadratic attention to cut inference cost, especially at long context. Released with open weights under Qwen's license in March 2025; now published as QRWKV-72B.
• Figma AI Figma · Design AI · 2024 — AI features in Figma. Auto-layout, content gen.
• Figure 02 Figure AI · Robotics / Physical AI · 2025 — Helix AI S1 subsystem ("The Motor Cortex," visuomotor control at 200 Hz). Figure 02 launched August 2024.
• Figure 03 Figure AI · Robotics / Physical AI · 2025 — Helix AI S2 subsystem ("The Brain," high-level reasoning at 7-10 Hz). Figure 03 launched October 2025.
• Fireflies.ai Fireflies · Transcription · 2023 — AI meeting recorder and summarizer.
• Fish Speech 1.5 Fish Audio · Speech Synthesis · 2024 — Fish Audio's Fish Speech 1.5, a text-to-speech model trained on more than 1 million hours of multilingual audio, led by English and Chinese (over 300K hours each) and Japanese (over 100K), with about 20K hours each for German, French, Spanish, Korean, Arabic and Russian and smaller amounts for Dutch, Italian, Polish and Portuguese; the weights are free for non-commercial use under CC BY-NC-SA 4.0.
• Fliki Fliki · Video Generation · 2022 — Text-to-video with AI voices.
• Fotor AI Fotor · Image Editing · 2023 — AI photo editing and enhancement.
• Fractal Analytics Fathom-R1-14B Fractal Analytics · LLM · 2025 — 14B math-reasoning model built from a DeepSeek-R1-distilled 14B model with a $499 training recipe, which Fractal says reaches o4-mini-level math reasoning within a 16K context. MIT license.
• Framer AI Framer · Web Design · 2023 — AI website generation and design.
• Fugaku-LLM Fujitsu, Tokyo Institute of Technology · LLM · 2024 — 13-billion-parameter language model with enhanced Japanese ability, trained from scratch on the RIKEN supercomputer Fugaku by Tokyo Institute of Technology, Tohoku University, Fujitsu, RIKEN, Nagoya University, CyberAgent and Kotoba Technologies; weights released for research and commercial use.
• Gamma Gamma · Presentation AI · 2023 — AI slide deck creation.
• Genmo Mochi 1 Genmo · Video Generation · 2024 — Genmo's Mochi 1 preview, a 10B-parameter open video generation model built on Genmo's Asymmetric Diffusion Transformer (AsymmDiT) and trained from scratch, released under Apache 2.0 with its AsymmVAE video compressor; Genmo called it the largest openly released video generative model at launch and cites high-fidelity motion and strong prompt adherence.
• Whisper.cpp ggerganov (community) · Inference Engine · 2023 — whisper.cpp, a dependency-free C/C++ implementation for fast local inference of OpenAI's Whisper speech recognition model, with integer quantization, zero memory allocations at runtime, voice activity detection, a C-style API and support for Apple silicon (Metal and Core ML), x86, NVIDIA, AMD and other accelerators; it runs on macOS, iOS, Android, Linux, Windows, WebAssembly and Raspberry Pi.
• GitHub Copilot GitHub (Microsoft) · Code Assistant · 2022 — GitHub Copilot, GitHub's AI coding assistant and agent platform: it offers completions, next-edit suggestions and multi-file agent mode in VS Code, Visual Studio, JetBrains IDEs, Xcode, Neovim and Eclipse, and on GitHub developers can assign issues to cloud agents (Copilot, Anthropic's Claude or OpenAI Codex), use Copilot code review, the Copilot app and the CLI, and choose among Claude, Gemini and GPT models. GitHub cites millions of individual users and tens of thousands of business customers.
• Glif Glif · AI Platform · 2024 — Visual AI workflow builder.
• ALBERT Google · LLM · 2019 — Google's ALBERT (A Lite BERT), which uses two parameter-reduction techniques to cut memory use and speed up BERT training, plus a self-supervised loss for inter-sentence coherence; its best model set state-of-the-art results on GLUE, RACE and SQuAD with fewer parameters than BERT-large, and the code and pretrained models are on GitHub.
• Bard Google · Chat Interface · 2023 — Google's Bard (opened March 21, 2023 in the U.S. and U.K.), an early experiment in chatting with generative AI, first powered by a lightweight, optimized version of LaMDA and designed as a complement to Google Search, with a 'Google it' button, multiple response drafts and guardrails such as caps on dialogue length; it moved to a fine-tuned Gemini Pro on December 6, 2023 and was renamed Gemini, alongside the launch of Gemini Advanced with Ultra 1.0, on February 8, 2024.
• BERT Google · LLM · 2018 — BERT (Bidirectional Encoder Representations from Transformers), Google's 2018 language model, the first deeply bidirectional one pre-trained only on plain text: it learns by predicting masked words from context on both sides and whether one sentence follows another, and fine-tunes with one added output layer. It set records on 11 NLP tasks, including 80.5% on GLUE and 93.2 F1 on SQuAD v1.1; Google open-sourced it that November and in 2019 used it to understand 1 in 10 US English searches.
• BERT-Large Google · Embeddings · 2018 — BERT-Large, the 340M-parameter version of Google's BERT (Bidirectional Encoder Representations from Transformers, October 2018), which pre-trains deep bidirectional representations from unlabeled text by conditioning on both left and right context in all layers, so one added output layer fine-tunes it for tasks such as question answering and language inference; it set new state-of-the-art results on eleven NLP tasks, including 80.5% on GLUE, 86.7% on MultiNLI and 93.2 F1 on SQuAD v1.1.
• ByT5 Google · LLM · 2021 — Google's ByT5, token-free T5-architecture models that read and write raw bytes instead of word or subword tokens, so they handle text in any language out of the box; Google reports they are competitive with token-level models, much more robust to noise and better on spelling- and pronunciation-sensitive tasks, and released the models, code and data.
• CoAtNet Google · Vision/CV · 2021 — Google's CoAtNet (June 2021), a family of hybrid image models that unify depthwise convolution and self-attention through relative attention and stack convolution and attention layers vertically for better generalization, capacity and efficiency; it reached 86.0% ImageNet top-1 accuracy with no extra data, 88.56% after pre-training on ImageNet-21K's 13M images, matching ViT-Huge pre-trained on 300M JFT images with 23x less data, and a record 90.88% when scaled up with JFT-3B.
• Deep Dream Generator Google · Image Generation · 2015 — Neural style transfer and deep dream visualization.
• EfficientNet Google · Vision/CV · 2019 — Google's EfficientNet family of convolutional networks, built from a neural-architecture-searched baseline scaled up with a compound coefficient that balances depth, width and resolution; EfficientNet-B7 reached a state-of-the-art 84.3% ImageNet top-1 accuracy while being 8.4x smaller and 6.1x faster at inference than the best previous ConvNet, and the models transfer well to CIFAR-100, Flowers and other datasets.
• EfficientNetV2-XL Google · Vision/CV · 2021 — Google's EfficientNetV2 (April 2021), a family of convolutional networks found by training-aware neural architecture search and scaling, with new ops such as Fused-MBConv, that train much faster than prior models while being up to 6.8x smaller; progressive learning grows image size while adapting regularization, and the largest model, EfficientNetV2-XL, pre-trained on ImageNet-21k reaches 87.3% ImageNet top-1 accuracy, 2.0% above ViT while training 5x to 11x faster on the same resources.
• ELECTRA Google · LLM · 2020 — Google's ELECTRA, which pretrains text encoders with replaced-token detection: a small generator swaps in plausible tokens and the model learns to spot which tokens were replaced, a task defined over every input token; Google reports it substantially beats BERT at the same size, data and compute, a one-GPU, four-day model outperformed GPT on GLUE, and at scale it matches RoBERTa and XLNet with under a quarter of their compute.
• FLAN 137B Google · LLM · 2022 — Google Research's FLAN (Fine-tuned LAnguage Net), built by instruction tuning: fine-tuning LaMDA-PT, a 137B-parameter pretrained language model, on over 60 NLP tasks phrased as natural-language instructions. On unseen task types it improved substantially on the untuned model, surpassed zero-shot 175B GPT-3 on 20 of 25 tasks, and beat few-shot GPT-3 by a large margin on ANLI, RTE, BoolQ, AI2-ARC, OpenbookQA and StoryCloze.
• FLAN-T5 Google · LLM · 2022 — Google's Flan-T5 checkpoints: T5 models instruction-finetuned with the scaled-up Flan recipe (more tasks and chain-of-thought data) from the paper 'Scaling Instruction-Finetuned Language Models'; Google reports strong few-shot performance even compared with much larger models such as PaLM 62B.
• FLAN-UL2 Google · LLM · 2023 — Google's Flan-UL2, a 20B encoder-decoder model with the T5 architecture and UL2 configuration, instruction-tuned on the Flan collection; it uses a 2,048-token receptive field (up from UL2's 512) and no longer needs UL2's mode tokens.
• GLaM Google · LLM · 2021 — Google's GLaM (Generalist Language Model, December 2021), a sparsely activated mixture-of-experts language model with 1.2 trillion parameters, about 7x GPT-3, in which a gating network sends each token to two experts so only 97B parameters (8%) are active; across 29 benchmarks it matched or beat GPT-3 on almost 80% of zero-shot and almost 90% of one-shot tasks while using a third of GPT-3's training energy and half its inference compute.
• Google AI Overview Google · Search AI · 2024 — Google Search's AI Overviews, AI-generated summaries at the top of results that give a quick overview of a topic with links to learn more; after being used billions of times in the Search Labs experiment, they began rolling out to everyone in the U.S. on May 14, 2024, powered by a custom Gemini model for Search with multi-step reasoning, planning and multimodality. Google said they would reach over a billion people by the end of 2024 and that links in them get more clicks than standard listings.
• Google Assistant (Gemini) Google · Virtual Assistant · 2024 — Gemini as the successor to Google Assistant: on March 14, 2025, with millions of users already switched, Google said it would upgrade mobile users from Assistant to Gemini, make the classic Assistant unavailable on most mobile devices later that year, and move tablets, cars, headphones and watches to Gemini, which adds Gemini Live conversations and Deep Research. On speakers and displays, Gemini for Home, announced August 20, 2025, replaces Assistant, starting with early access in October.
• Google Gemini for Workspace Google · Productivity AI · 2024 — Gemini for Google Workspace, Google's generative AI in Gmail, Docs, Sheets, Slides, Meet and other Workspace apps, which succeeded Duet AI on February 21, 2024: it drafts and summarizes email and documents, fills in Sheets, makes images in Slides, and takes notes and translates captions in Meet. It launched as Gemini Business and Enterprise add-ons at $20 and $30 per user a month; since January 2025 it is included in Workspace Business and Enterprise plans, with the Gemini app and NotebookLM.
• Google Translate (Neural) Google · Translation · 2016 — Google Neural Machine Translation (GNMT), announced September 27, 2016, which, unlike the phrase-based system Google Translate had used since launch, translates whole sentences as a unit with recurrent neural networks and attention; in side-by-side ratings by bilingual human raters it cut translation errors by 55% to 85% on several major language pairs, and it launched in production for Chinese to English, about 18 million translations a day, running on TensorFlow and Tensor Processing Units.
• GoogLeNet/Inception v1 Google · Vision/CV · 2014 — Google's GoogLeNet, the 22-layer incarnation of the Inception convolutional architecture, which increased network depth and width while keeping the compute budget constant; it set the new state of the art for classification and detection in the ImageNet Large-Scale Visual Recognition Challenge 2014 (ILSVRC 2014).
• Inception v3 Google · Vision/CV · 2015 — Google's Inception v3, from 'Rethinking the Inception Architecture for Computer Vision', which scales convolutional networks efficiently with factorized convolutions and aggressive regularization; a single network with under 25 million parameters and 5 billion multiply-adds per inference scored 21.2% top-1 and 5.6% top-5 error on ILSVRC 2012 validation, and a four-model ensemble reached 3.5% top-5 error (3.6% on test).
• LaMDA Google · LLM · 2021 — Google's LaMDA (Language Model for Dialogue Applications, announced May 18, 2021), a family of Transformer language models specialized for open-ended dialogue, with up to 137B parameters pre-trained on 1.56 trillion words of public dialogue data and web text; fine-tuning on annotated data and letting it consult external tools such as a retrieval system, translator and calculator improved its safety and factual grounding, and a lightweight version first powered Bard in 2023.
• Med-PaLM Google · Medical LLM · 2022 — Google Research's Med-PaLM, a medical question-answering model made by aligning Flan-PaLM (the instruction-tuned 540B-parameter PaLM) to medicine with instruction prompt tuning; preprinted in December 2022 and published in Nature in July 2023, it was the first AI system to pass (above 60%) US Medical Licensing Exam-style questions, though its authors found it still inferior to clinicians.
• Med-PaLM 2 Google · Medical LLM · 2023 — Google Research's Med-PaLM 2, introduced in March 2023, which combined PaLM 2 with medical fine-tuning and prompting strategies including ensemble refinement to score 86.5% on MedQA US Medical Licensing Exam-style questions, the first expert-level result and over 19% better than Med-PaLM; physicians preferred its long-form answers to physicians' own on eight of nine axes, and it was one of the research models behind Google Cloud's MedLM.
• Minerva Google · LLM · 2022 — Google Research's Minerva (June 2022), a language model for mathematical and scientific questions built on PaLM and further trained on a 118GB dataset of arXiv papers and web pages with mathematical notation; with chain-of-thought prompting and majority voting it set state-of-the-art results on technical benchmarks such as MATH and MMLU-STEM without external tools, and it correctly answered nearly a third of over 200 undergraduate-level science problems.
• MobileNetV3 Google · Vision/CV · 2019 — Google's MobileNetV3, convolutional networks tuned for mobile phone CPUs with hardware-aware neural architecture search, the NetAdapt algorithm and new architecture changes, released as MobileNetV3-Large and -Small; versus MobileNetV2, Large is 3.2% more accurate on ImageNet with 15% lower latency and Small 4.6% more accurate with 5% lower latency, and a new LR-ASPP decoder handles segmentation.
• mT5 Google · LLM · 2020 — Google's mT5, a multilingual version of the T5 text-to-text transformer pretrained on a new Common Crawl-based dataset covering 101 languages; Google reported state-of-the-art results on many multilingual benchmarks, described a technique that prevents 'accidental translation' in zero-shot use, and released all code and checkpoints.
• NASNet Google · Vision/CV · 2017 — Google's NASNet, an image-classification architecture found by neural architecture search: the best convolutional cell was searched on CIFAR-10 and stacked to build ImageNet models; it reached 82.7% top-1 on ImageNet, 1.2% better than the best human-designed architectures with 28% less computation, and its features lifted COCO detection to 43.1% mAP with Faster R-CNN.
• OWL-ViT Google · Vision/CV · 2022 — Google's OWL-ViT, an open-vocabulary object detector that adapts a contrastively pretrained image-text Vision Transformer to detection with minimal changes and end-to-end fine-tuning; it finds objects from free-text queries (zero-shot) or from an example image (one-shot), and its authors found more image-level pretraining and larger models consistently improved detection.
• OWLv2 Google · Vision/CV · 2023 — Google's OWLv2 open-vocabulary object detector and its OWL-ST self-training recipe, which uses an existing detector to pseudo-label web image-text pairs so detection training can scale past 1 billion examples; with an L/14 model, OWL-ST raised AP on LVIS rare classes, which the model never saw human box labels for, from 31.2% to 44.6%.
• PaLM Google · LLM · 2022 — Google's Pathways Language Model, PaLM (April 4, 2022), a 540-billion-parameter dense Transformer trained on 6,144 TPU v4 chips across two Pods with the Pathways system, on web documents, books, Wikipedia, conversations and GitHub code; it set few-shot state-of-the-art results on hundreds of benchmarks, beat average human performance on BIG-bench, and with chain-of-thought prompting solved 58% of GSM8K grade-school math problems, above the prior best of 55%. PaLM 2 succeeded it in May 2023.
• PaLM (540B) Google · LLM · 2022 — Google's Pathways Language Model, PaLM (April 4, 2022), a 540-billion-parameter dense Transformer trained on 6,144 TPU v4 chips across two Pods with the Pathways system, on web documents, books, Wikipedia, conversations and GitHub code; it set few-shot state-of-the-art results on hundreds of benchmarks, beat average human performance on BIG-bench, and with chain-of-thought prompting solved 58% of GSM8K grade-school math problems, above the prior best of 55%. PaLM 2 succeeded it in May 2023.
• PaLM 2 Google · LLM · 2023 — Google's PaLM 2 (May 10, 2023), its next-generation language model with stronger multilingual, reasoning and coding skills: trained more heavily on text in over 100 languages, it passes advanced language proficiency exams at the 'mastery' level, and pre-training on public source code helps it in Python and JavaScript as well as Prolog, Fortran and Verilog. It came in four sizes (Gecko, Otter, Bison and Unicorn), powered Bard's expansion to new languages, and spawned Med-PaLM 2 and Sec-PaLM.
• PaLM-E Google · Multimodal LLM · 2023 — Google's PaLM-E (March 2023), an embodied multimodal language model that feeds continuous sensor data such as images and robot state estimates into a pre-trained language model as 'multimodal sentences', trained end to end for robotic manipulation planning, visual question answering and captioning; the largest version, PaLM-E-562B, transfers knowledge across language, vision and robotics data, set a state-of-the-art result on OK-VQA, and keeps its general language skills.
• Photomath Google · Education AI · 2023 — Photomath, the math-learning app created by Damir Sabol that joined Google in 2023: users scan a problem with a phone camera, anything from arithmetic to calculus, and a neural network on Photomath's cloud servers reads it before an expert-system solver returns free step-by-step solutions. Photomath Plus, an in-app subscription billed monthly, every six months or yearly, adds more than 400 AI-powered animated tutorials, deeper explanations and contextual hints.
• Seq2Seq LSTM Google · LLM · 2014 — Google's sequence-to-sequence learning paper (Sutskever, Vinyals and Le, September 2014), which used one multilayer LSTM to encode an input sentence into a fixed-size vector and another deep LSTM to decode the output; on WMT'14 English-to-French it scored 34.8 BLEU, above a phrase-based system's 33.3, and 36.5 when reranking that system's top 1,000 outputs, handled long sentences well, and improved markedly when source sentences were reversed.
• Socratic (Google) Google · Education AI · 2022 — Socratic by Google, a homework-help app for high school and university students that Google acquired in 2018 and that by August 2019 used AI to help: students photograph, type or speak a question, and Socratic identifies the underlying concepts and points them to the most relevant videos, explanations and web resources, plus subject guides on over 1,000 topics. The socratic.org site now redirects to Google Lens, which offers step-by-step homework help.
• SwitchTransformer Google · LLM · 2021 — Google's Switch Transformer, a sparsely activated mixture-of-experts Transformer with a simplified routing algorithm that lowers communication and compute costs and can be trained in bfloat16; versions based on T5-Base and T5-Large pretrained up to 7x faster with the same compute, and trillion-parameter models trained on the C4 corpus reached a 4x speedup over T5-XXL.
• T5 Google · LLM · 2019 — Google's T5 (Text-to-Text Transfer Transformer, October 2019), a framework that casts every text task, from question answering and classification to summarization, as feeding text in and generating text out; Google's systematic study of pre-training objectives, architectures and data sets, combined with scale and its new Colossal Clean Crawled Corpus (C4), set state-of-the-art results on GLUE, SuperGLUE, SQuAD and CNN/Daily Mail, and the data set, models and code were released.
• T5-11B Google · LLM · 2019 — T5-11B, the largest model in Google's T5 (Text-to-Text Transfer Transformer) family from October 2019, with 11 billion parameters; combining the best methods from Google's systematic transfer-learning study with scale and the Colossal Clean Crawled Corpus (C4), it set state-of-the-art results on GLUE, SuperGLUE, SQuAD and CNN/Daily Mail and reached a near-human score on SuperGLUE, a benchmark designed to be hard for machines but easy for humans.
• T5-3B Google · LLM · 2019 — Google's T5-3B, the 3-billion-parameter version of T5 (Text-to-Text Transfer Transformer), which casts every language task in a text-to-text format; built on Google's systematic study of pretraining objectives, architectures and datasets plus the new Colossal Clean Crawled Corpus (C4), the T5 family set state-of-the-art results in summarization, question answering and text classification, with data, models and code released.
• TimesFM-ICF Google · LLM · 2025 — Version of Google's TimesFM forecasting model with in-context fine-tuning: at inference it can be prompted with examples from related time series to adapt its forecasts to the target domain.
• Transformer Google · LLM · 2017 — The Transformer, the neural network architecture introduced in Google researchers' June 2017 paper Attention Is All You Need: it dropped recurrence and convolutions and relied solely on attention, which made training more parallelizable and much faster. It scored 28.4 BLEU on WMT 2014 English-German translation, over 2 BLEU above prior bests, ensembles included, set a single-model record of 41.8 BLEU on English-French after 3.5 days on eight GPUs, and also handled English constituency parsing.
• Transformer (original) Google · Architecture · 2017 — The Transformer from Google's 2017 paper 'Attention Is All You Need', a sequence-to-sequence architecture based solely on attention, with no recurrence or convolutions; it set new machine-translation records (28.4 BLEU on WMT 2014 English-German, 41.8 on English-French after 3.5 days on eight GPUs) while being more parallelizable and far cheaper to train, and also worked for English constituency parsing.
• UL2 Google · LLM · 2022 — Google's UL2 20B, pretrained with Mixture-of-Denoisers, an objective that combines several pretraining paradigms, plus mode switching tied to downstream tasks; Google reports state-of-the-art results on 50 supervised NLP tasks, zero-shot SuperGLUE above 175B GPT-3 and strong chain-of-thought reasoning, and released T5X checkpoints for UL2 20B and Flan-UL2 20B.
• ViT-22B Google · Vision/CV · 2023 — Google's ViT-22B, a 22-billion-parameter Vision Transformer trained with a recipe for efficient, stable training at that scale (the largest dense ViT before it had 4B parameters); Google reports downstream performance that keeps improving with scale, a better fairness-performance tradeoff, state-of-the-art alignment with human shape/texture bias and improved robustness.
• ViT-Base Google · Vision/CV · 2020 — Google's Vision Transformer (ViT), introduced in 'An Image is Worth 16x16 Words' (October 2020), which applies a standard Transformer directly to sequences of image patches instead of relying on convolutional networks; pre-trained on large datasets and transferred to benchmarks such as ImageNet, CIFAR-100 and VTAB, it matched or beat state-of-the-art convolutional networks while needing substantially less compute to train. ViT-Base is the 86M-parameter base size of the family.
• Word2Vec (large) Google · Embeddings · 2013 — Google's word2vec research (Mikolov, Chen, Corrado and Dean, 2013), which proposed two efficient architectures, continuous bag-of-words (CBOW) and continuous skip-gram, for learning vector representations of words from very large data sets; they beat earlier neural-network methods on word similarity at much lower computational cost, learning high-quality vectors from a 1.6-billion-word data set in under a day, with state-of-the-art results on syntactic and semantic similarity tests.
• Xception Google · Vision/CV · 2016 — Google's Xception, a convolutional network that replaces Inception modules with depthwise separable convolutions; with the same number of parameters as Inception V3, it slightly outperforms Inception V3 on ImageNet and significantly outperforms it on a 350-million-image, 17,000-class dataset, gains the paper attributes to more efficient use of parameters.
• XLNet Google · LLM · 2019 — XLNet (June 2019), from Carnegie Mellon University and the Google AI Brain Team, an autoregressive pretraining method that learns bidirectional context by maximizing likelihood over all permutations of the factorization order, avoiding BERT's masked inputs and pretrain-finetune mismatch, and that borrows from Transformer-XL; under comparable settings it outperformed BERT on 20 tasks, often by a large margin, including question answering, inference, sentiment analysis and document ranking.
• ViT-G/14 Google Brain · Vision/CV · 2022 — Google Brain's ViT-G/14, the two-billion-parameter Vision Transformer from the 'Scaling Vision Transformers' study, which mapped how ViT error rates scale with data and compute and refined ViT training to cut memory use; it set a then state-of-the-art 90.45% top-1 accuracy on ImageNet and reached 84.86% with only 10 examples per class.
• ViT-Huge/14 Google Brain · Vision/CV · 2020 — ViT-Huge/14, a model from Google Brain's original Vision Transformer paper ('An Image is Worth 16x16 Words'), which showed that a pure transformer applied to sequences of image patches, pretrained on large datasets, matches or beats state-of-the-art convolutional networks on ImageNet, CIFAR-100 and VTAB while needing substantially less compute to train.
• Agentic-Tx Google DeepMind · Code Agent · 2025 — Google's Agentic-Tx, a research agent system for therapeutics in which a Gemini model works through multi-step drug-discovery questions using 18 tools, including TxGemma, PubMed, Wikipedia and web search, and molecule, gene and protein tools; Google first showed it on Gemini 2.0 Pro when it released TxGemma, and its paper, running it on Gemini 2.5, reports a 52.3% relative gain over o3-mini (high) on Humanity's Last Exam chemistry and biology and 26.7% on GPQA chemistry.
• AlphaFold 3 Google DeepMind · Science/Biology · 2024 — Google DeepMind and Isomorphic Labs' AlphaFold 3 (May 8, 2024), which predicts the joint 3D structure of proteins, DNA, RNA, ligands and chemical modifications using an improved Evoformer and a diffusion network; Google reports at least 50% better predictions than existing methods for protein interactions with other molecules and 50% higher accuracy than the best traditional methods on PoseBusters. AlphaFold Server is free for research, and academic weights followed in November 2024.
• Atlas Google DeepMind · LLM · 2025
• CodeGemma Google DeepMind · Code LLM · 2024 — Google's CodeGemma (April 9, 2024), open code models built on Gemma and further trained on 500 billion to 1 trillion more tokens of mostly English math data and synthetic code: a 2B pre-trained model for fast fill-in-the-middle code completion, and 7B pre-trained and instruction-tuned models for code generation and chat that keep strong language understanding and math reasoning. The 7B instruction-tuned v1.1, released May 3, 2024, scores 60.4 on HumanEval.
• DolphinGemma Google DeepMind · LLM · 2025 — Audio model of about 400M parameters from Google, built with the Wild Dolphin Project and Georgia Tech, that learns the structure of dolphin vocalizations and predicts the next sounds in a sequence. Small enough to run on the Pixel phones researchers use in the field.
• EmbeddingGemma Google DeepMind · Embeddings · 2025 — Google's EmbeddingGemma (September 4, 2025), a 308M-parameter open text embedding model built on Gemma 3 for on-device search and retrieval-augmented generation; trained on 100+ languages with a 2K-token context, it outputs 768-dimension vectors that can be shortened to 512, 256 or 128 via Matryoshka learning, runs in under 200MB of RAM when quantized, and was the top open multilingual embedding model under 500M parameters on MTEB at launch. EmbeddingGemma 2 (740M) followed in October 2026.
• Flow Google DeepMind · Video Production · 2025 — Google's Flow (May 20, 2025), an AI filmmaking tool custom-designed for Google's Veo, Imagen and Gemini models: users describe shots in everyday language, create or upload 'ingredients' such as characters and reuse them consistently across clips, and use camera controls, Scenebuilder and asset management. It launched for Google AI Pro and Ultra subscribers in the U.S., Veo 3.1 later added sound across its features, and by October 2025 people had made over 275 million videos in it.
• Gecko Google DeepMind · Embeddings · 2024 — Google DeepMind's Gecko, a compact text embedding model trained by distilling knowledge from large language models: an LLM generates synthetic query-passage pairs, then relabels positive and hard-negative passages; on the MTEB benchmark the 256-dimension version beat all existing 768-dimension entries, and the 768-dimension version scored 66.31, competitive with 7x larger models.
• Gemini 1.0 Google DeepMind · Multimodal LLM · 2023 — Google DeepMind's Gemini 1.0 (December 6, 2023), Google's first Gemini model family, built to be natively multimodal by pre-training on different modalities from the start, in three sizes: Ultra for highly complex tasks, Pro for a wide range of tasks and Nano for on-device use; Gemini Pro powered Bard from launch day and reached developers on December 13, Nano ran first on the Pixel 8 Pro, and Ultra beat prior state-of-the-art results on 30 of 32 benchmarks and scored 90.0% on MMLU.
• Gemini 1.0 Ultra Google DeepMind · Multimodal LLM · 2023 — Google DeepMind's Gemini Ultra (Ultra 1.0), the largest and most capable model in the December 6, 2023 Gemini 1.0 family, built for highly complex tasks; Google reported state-of-the-art results on 30 of 32 widely used benchmarks, 90.0% on MMLU (the first model to beat human experts there) and 59.4% on MMMU. After further safety testing it reached consumers on February 8, 2024 through Gemini Advanced, part of the $19.99-a-month Google One AI Premium plan.
• Gemini 1.5 Google DeepMind · Multimodal LLM · 2024 — Google's Gemini 1.5 family of multimodal models with context windows of millions of tokens, able to recall and reason over long documents and hours of video and audio: Gemini 1.5 Pro (February 2024, a mixture-of-experts model on par with Gemini 1.0 Ultra) and the lighter, distilled Gemini 1.5 Flash (May 2024); Google reports near-perfect retrieval (over 99%) up to at least 10 million tokens in research tests and 26% to 75% time savings for professionals across 10 job categories.
• Gemini 1.5 Flash Google DeepMind · Multimodal LLM · 2024 — Google's Gemini 1.5 Flash, introduced at I/O on May 14, 2024 as a lighter-weight, faster and cheaper-to-serve sibling of Gemini 1.5 Pro, distilled from 1.5 Pro while keeping its multimodal reasoning and 1-million-token context; Google aimed it at high-volume, high-frequency tasks such as summarization, chat, image and video captioning and data extraction from long documents and tables, and it was the fastest Gemini model in the API at launch.
• Gemini 1.5 Pro Google DeepMind · Multimodal LLM · 2024 — Google's Gemini 1.5 Pro, a mid-size multimodal mixture-of-experts model introduced February 15, 2024 that performed at about the level of Gemini 1.0 Ultra; it launched with a 128K-token context and a 1-million-token preview, enough for an hour of video, 11 hours of audio or over 700,000 words, found planted facts 99% of the time at 1 million tokens, and opened a 2-million-token context window to all developers in June 2024.
• Gemini 2.0 Google DeepMind · Multimodal LLM · 2024 — Google DeepMind's Gemini 2.0 (December 11, 2024), the model generation Google built for the 'agentic era', launched with an experimental 2.0 Flash that Google said beat 1.5 Pro on key benchmarks at twice the speed, with native tool use such as Google Search and code execution, image and speech output for early-access partners, and a Multimodal Live API for real-time audio and video. 2.0 Flash became generally available on February 5, 2025, with experimental 2.0 Pro and a Flash-Lite preview.
• Gemini 2.0 Flash exp Google DeepMind · Multimodal LLM · 2024 — Google's experimental Gemini 2.0 Flash (December 11, 2024), the first Gemini 2.0 model, opened to developers in Google AI Studio and Vertex AI and to all Gemini app users; Google said it outperformed 1.5 Pro on key benchmarks at twice the speed, took image, video and audio input, natively called tools such as Google Search and code execution, and could generate images and steerable multilingual speech for early-access partners. A generally available 2.0 Flash followed on February 5, 2025.
• Gemini 2.0 Pro Google DeepMind · Multimodal LLM · 2025 — Google's Gemini 2.0 Pro (February 5, 2025), released as an experimental model in Google AI Studio, Vertex AI and the Gemini app for Gemini Advanced users; Google called it its best model yet for coding and complex prompts, with better world-knowledge reasoning than any model it had released, its largest context window at 2 million tokens, and tool calls to Google Search and code execution. It launched alongside a generally available 2.0 Flash and a 2.0 Flash-Lite preview.
• Gemini 2.5 Google DeepMind · Multimodal LLM · 2025 — Google DeepMind's Gemini 2.5 (March 25, 2025), a family of thinking models that reason through their thoughts before responding; the first release, an experimental 2.5 Pro, debuted at #1 on LMArena by a significant margin, led GPQA and AIME 2025 without test-time techniques, scored 18.8% on Humanity's Last Exam without tools and 63.8% on SWE-Bench Verified with a custom agent, and shipped with a 1M-token context. 2.5 Flash followed in April, and both became generally available on June 17, 2025.
• Gemini 2.5 Flash Google DeepMind · Multimodal LLM · 2025 — Google's Gemini 2.5 Flash (preview April 17, 2025; generally available June 17, 2025), the price-performance model of the Gemini 2.5 family and Google's first fully hybrid reasoning model: developers can turn thinking on or off and cap it with a thinking budget of up to 24,576 tokens. It takes text, image, video and audio input with a 1,048,576-token input limit and 65,536-token output, and Google said it ranked second only to 2.5 Pro on LMArena's Hard Prompts at launch.
• Gemini 2.5 Flash Preview Google DeepMind · Multimodal LLM · 2025 — Gemini 2.5 Flash Preview, the updated Gemini 2.5 Flash Google released at I/O on May 20, 2025 as its most efficient workhorse model: it improved on reasoning, multimodality, code and long-context benchmarks while using 20-30% fewer tokens, added thought summaries in the API and kept thinking budgets for cost control. It reached Google AI Studio, Vertex AI and the Gemini app; the stable model followed June 17, 2025, and this preview shut down November 18, 2025.
• Gemini 2.5 Flash-Lite Google DeepMind · Multimodal LLM · 2025 — Fastest, lowest-cost Gemini 2.5 model, generally available Jul 22, 2025 after a June preview: 1M-token context, optional thinking budgets, and native tools (Grounding with Google Search, Code Execution, URL Context). Google reports lower latency than 2.0 Flash-Lite and 2.0 Flash.
• Gemini 2.5 Pro Google DeepMind · Multimodal LLM · 2025 — Google's Gemini 2.5 Pro (experimental March 25, 2025; generally available June 17, 2025), a thinking model that reasons before responding and takes text, images, video, audio and PDFs, with a 1,048,576-token input limit, 65,536-token output and a January 2025 knowledge cutoff; at launch it debuted at #1 on LMArena by a significant margin, scored 18.8% on Humanity's Last Exam without tools and 63.8% on SWE-Bench Verified with a custom agent setup.
• Gemini 2.5 Pro 06-05 Google DeepMind · Multimodal LLM · 2025 — Gemini 2.5 Pro 06-05, the upgraded Gemini 2.5 Pro preview Google released June 5, 2025, ahead of the stable model on June 17: it gained 24 points on LMArena to lead at 1470 Elo and 35 points on WebDevArena to lead at 1443, led on the Aider Polyglot coding benchmark, wrote more creative, better-formatted responses and added thinking budgets for cost and latency control. It ran in the Gemini app and the Gemini API; the preview shut down December 2, 2025.
• Gemini 3 Google DeepMind · Multimodal LLM · 2025 — Google's Gemini 3 (November 18, 2025), launched with Gemini 3 Pro in preview across the Gemini app, AI Mode in Search (Gemini's first day-one launch in Search), AI Studio, Vertex AI and Google's new Antigravity agentic development platform; Gemini 3 Pro topped LMArena at 1501 Elo and scored 37.5% on Humanity's Last Exam without tools, 91.9% on GPQA Diamond and 76.2% on SWE-bench Verified, with a 1M-token context. Google also announced a Deep Think mode for Google AI Ultra subscribers.
• Gemini 3 Deep Think Google DeepMind · Reasoning LLM · 2026 — Gemini 3 Deep Think, the Gemini 3 version of Google's specialized reasoning mode, which uses parallel reasoning to explore several hypotheses at once. Its major February 12, 2026 upgrade, built with scientists, scored 48.4% on Humanity's Last Exam without tools, 84.6% on ARC-AGI-2 and 3455 Elo on Codeforces, reached gold-medal level on IMO 2025 and the written 2025 Physics and Chemistry Olympiads, and opened Gemini API early access; it is in the Gemini app for Google AI Ultra subscribers.
• Gemini 3 Flash Google DeepMind · Multimodal LLM · 2025 — Google's Gemini 3 Flash (December 17, 2025), which pairs Gemini 3's Pro-grade reasoning with Flash speed and cost and replaced 2.5 Flash as the default model in the Gemini app, also becoming the default in AI Mode in Search; it scored 90.4% on GPQA Diamond, 33.7% on Humanity's Last Exam without tools, 81.2% on MMMU-Pro and 78% on SWE-bench Verified, above Gemini 3 Pro, and Google says it beats 2.5 Pro while running 3x faster and using 30% fewer tokens, at $0.50/$3 per million tokens.
• Gemini 3 Pro Google DeepMind · Multimodal LLM · 2025 — Google's Gemini 3 Pro (preview November 18, 2025), the first Gemini 3 model, which topped LMArena at 1501 Elo and WebDev Arena at 1487 and scored 37.5% on Humanity's Last Exam without tools, 91.9% on GPQA Diamond, 23.4% on MathArena Apex, 81% on MMMU-Pro, 76.2% on SWE-bench Verified and 54.2% on Terminal-Bench 2.0, with a 1M-token context; its Deep Think mode reached 41.0% on Humanity's Last Exam and 45.1% on ARC-AGI-2. Gemini 3.1 Pro has since succeeded it.
• Gemini 3.1 Flash-Lite Google DeepMind · Multimodal LLM · 2026 — Gemini 3.1 Flash-Lite, Google's fastest and most cost-efficient Gemini 3 series model, built on Gemini 3 Pro for high-volume, latency-sensitive work like translation, content moderation and classification. Previewed March 3, 2026 at $0.25/$1.50 per million input/output tokens, it reached its first answer token 2.5 times faster than Gemini 2.5 Flash, scored 86.9% on GPQA Diamond and has thinking levels and a 1M-token context. The stable model came May 7, 2026 and shuts down May 7, 2027.
• Gemini 3.1 Pro Google DeepMind · Multimodal LLM · 2026 — Google's Gemini 3.1 Pro, released in preview on February 19, 2026 as the upgraded core reasoning model of the Gemini 3 series and the intelligence behind the previous week's Gemini 3 Deep Think update; it scored a verified 77.1% on ARC-AGI-2, more than double Gemini 3 Pro, and rolled out to developers (Gemini API, Gemini CLI, Antigravity, Android Studio), enterprises (Vertex AI, Gemini Enterprise) and consumers in the Gemini app and NotebookLM.
• Gemini 3.5 Flash Google DeepMind · Multimodal LLM · 2026 — Gemini 3.5 Flash (May 19, 2026), the first model in Google's Gemini 3.5 series and its strongest agentic and coding model at launch: built on Gemini 3 Flash, it beat Gemini 3.1 Pro on Terminal-bench 2.1 (76.2% vs 70.3%), MCP Atlas (83.6%) and GDPval-AA (1656 Elo), ran four times faster than other frontier models in output speed and takes text, image, audio and video in a 1M-token context. It became the Gemini app's default and powers the Gemini Spark agent; API calls now go to Gemini 3.6 Flash.
• Gemini 3.5 Flash Cyber Google DeepMind · LLM · 2026 — Google DeepMind's Gemini 3.5 Flash Cyber (July 21, 2026), a lightweight cybersecurity model built on Gemini 3.5 Flash and fine-tuned to find, validate and patch software vulnerabilities, called many times by Google's CodeMender agent to scan large codebases cheaply; on the V8 JavaScript engine it found 55 unique confirmed issues versus 47 for 3.5 Flash and 36 for Claude Opus 4.6. Because of its dual-use risk, Google offers it only through a limited pilot for governments and trusted partners.
• Gemini 3.5 Flash-Lite Google DeepMind · LLM · 2026 — Google's Gemini 3.5 Flash-Lite (July 21, 2026), its fastest and most cost-efficient 3.5-class model, released with Gemini 3.6 Flash and 3.5 Flash Cyber for high-volume, latency-sensitive work and agents; built on 3.1 Flash-Lite, it takes text, images, audio and video with a 1M-token context and 64K output, has selectable thinking levels and built-in computer use, a March 2026 knowledge cutoff, and scores 54.2% on SWE-Bench Pro and 74.0% on OSWorld-Verified at $0.30/$2.50 per million tokens.
• Gemini 3.6 Flash Google DeepMind · LLM · 2026 — Launched as Google's default 'workhorse' Gemini: 17% lower output-token usage than 3.5 Flash (per Artificial Analysis), 1M-token context with 64K max output, 58.7% SWE-Bench Pro. Shipped with 3.5 Flash-Lite and gated 3.5 Flash Cyber. Google now calls it its previous-generation Flash model, after Gemini 3.8 Flash.
• Gemini 3.7 Flash Google DeepMind · LLM · 2026 — Google's Gemini 3.7 Flash (August 13, 2026), which Google called its most intelligent workhorse model yet for coding and agents, released three weeks after Gemini 3.6 Flash; it beat 3.6 Flash on DeepSWE v1.1 (65.3% vs 49.0%), FrontierCode 1.1 (43.6% vs 34.4%), WebDev Arena (1588 vs 1538 Elo) and GDP.pdf (34.0% vs 22.0%), takes text, images, video, audio and PDFs with a 1M-token context, and launched at half 3.6 Flash's price. Gemini API calls to it now route to 3.8 Flash.
• Gemini 3.8 Flash Google DeepMind · LLM · 2026 — Gemini 3.8 Flash (September 2, 2026), Google's most intelligent workhorse model, built on Gemini 3.7 Flash with gains in long-horizon coding, agentic tasks and multi-step reasoning: it scores 73.7% on DeepSWE v1.1, 89.4% on Terminal-bench 2.1 and 54.9% on HLE-Verified, takes text, image, audio and video input in a 1M-token context with 64K output, and has adjustable effort levels. API prices are $0.75/$3.75 per million input/output tokens through 2026, then $1.50/$7.50.
• Gemini 3.8 Flash Cyber Google DeepMind · LLM · 2026 — Gemini 3.8 Flash Cyber (September 2, 2026), Google's most capable cybersecurity model, sharing Gemini 3.8 Flash's core but with more permissive cyber safeguards, so it goes only to trusted defenders in Google's Fairwind Program. It reached frontier level at finding vulnerabilities on CyberGym, topped 70% on an internal test in 20 programming languages and scored 47.2% pass@1 on CWE-Bench patching; Chrome's security team found it wrote 2.6 times more correct patches than much larger models.
• Gemini App Google DeepMind · Chat Interface · 2024 — Google's Gemini app (February 8, 2024), launched when Google renamed Bard to Gemini and released an Android app and Gemini in the Google app on iOS, with Gemini Pro 1.0 free in over 40 languages and Gemini Advanced with Ultra 1.0 in the $19.99-a-month Google One AI Premium plan; it has since added Canvas and Audio Overviews, Deep Think for Ultra subscribers, Lyria 3 music and Nano Banana Pro images, and in December 2025 Gemini 3 Flash became its default model.
• Gemini Canvas Google DeepMind · AI Writing · 2025 — Google's Canvas in the Gemini app (March 18, 2025), an interactive space opened from the prompt bar where users write and edit documents or code with Gemini and see changes in real time: Gemini drafts, suggests edits and adjusts tone, length or formatting, exports documents to Google Docs, and turns coding ideas into web apps, Python scripts, games and simulations with live HTML and React previews. It rolled out globally to Gemini and Gemini Advanced users.
• Gemini CLI Google DeepMind · Code Agent · 2025 — Google's Gemini CLI (June 25, 2025), an open-source (Apache 2.0) AI agent that brings Gemini into the terminal for coding, debugging, content generation and research; it can ground prompts with Google Search, extend itself through the Model Context Protocol and run non-interactively in scripts, and it shares technology with Gemini Code Assist. At launch a personal Google account gave free access to Gemini 2.5 Pro, with its 1M-token context, at 60 requests a minute and 1,000 a day.
• Gemini Deep Think Google DeepMind · Reasoning LLM · 2025 — Google's Gemini Deep Think, a reasoning mode that uses parallel thinking to explore and combine several candidate answers with longer thinking time; an advanced version solved five of six problems at IMO 2025 for an officially certified gold-medal score of 35 of 42, working end to end in natural language, and a faster version reached Google AI Ultra subscribers in the Gemini app in August 2025. Gemini 3.1 Deep Think, built on Gemini 3.1 Pro, is the current version.
• Gemini Diffusion Google DeepMind · Multimodal LLM · 2025 — Experimental text diffusion language model from Google DeepMind that generates whole blocks of tokens by refining noise step by step, rather than one token at a time; DeepMind reports about 1,479 tokens per second. Available only as an experimental demo.
• Gemini Flash Google DeepMind · Multimodal LLM · 2024 — Gemini Flash, Google's line of fast, cost-efficient Gemini models. It began with Gemini 1.5 Flash, introduced May 14, 2024 as a lighter-weight model distilled from 1.5 Pro and optimized for high-volume, high-frequency tasks such as summarization, captioning and data extraction, with a 1M-token context window. The current model, Gemini 3.8 Flash, is Google's most intelligent workhorse model for coding and agents, taking text, image, audio, video and PDF input with 1M input and 64K output tokens.
• Gemini Nano Google DeepMind · On-device LLM · 2023 — Google's Gemini Nano, the most efficient model of the December 6, 2023 Gemini 1.0 family, built for on-device tasks and first run on the Pixel 8 Pro; it runs inside Android's AICore system service, so apps can offer generative AI without a network connection or sending data to the cloud, and developers reach it through ML Kit GenAI APIs for prompting with text or images, summarization, proofreading, rewriting, image description and speech recognition.
• Gemini Robotics Google DeepMind · Multimodal LLM · 2025 — Gemini Robotics (March 12, 2025), Google DeepMind's vision-language-action model built on Gemini 2.0 that adds physical actions as an output to control robots directly. DeepMind said it more than doubled rival VLA models' generalization scores, follows conversational commands in several languages, replans when objects move, and handles dexterous tasks like origami folding; trained mainly on ALOHA 2 robots, it also controlled Franka arms and could be adapted to Apptronik's Apollo humanoid.
• Gemini Robotics 1.5 Google DeepMind · Multimodal LLM · 2025 — Google DeepMind's most capable vision-language-action model: turns visual input and instructions into robot motor commands, thinks before acting and shows its process, and learns across robot embodiments; available to select partners.
• Gemini Robotics-ER Google DeepMind · Multimodal LLM · 2025 — Gemini Robotics-ER (March 12, 2025), Google DeepMind's embodied-reasoning vision-language model, based on Gemini 2.0 with the spatial understanding robots need, that roboticists can connect to their own low-level controllers. It improved sharply on Gemini 2.0 at pointing and 3D detection, could plan a grasp and a safe approach for an object, and, handling perception, state estimation, planning and code generation itself, succeeded 2-3 times as often as Gemini 2.0. It went to trusted testers.
• Gemini Robotics-ER 1.6 Google DeepMind · RL/Agents · 2026 — Gemini Robotics-ER 1.6 (April 14, 2026), Google DeepMind's reasoning-first robotics model, a vision-language model based on Gemini 3.0 Flash that acts as a robot's high-level planner, calling tools like Google Search or vision-language-action models. It beat ER 1.5 and Gemini 3.0 Flash at pointing, counting and success detection, even across camera views, and added instrument reading, built with Boston Dynamics for Spot. Its API preview shut down August 31, 2026; Robotics-ER 2 replaced it.
• Gemini Ultra Google DeepMind · Multimodal LLM · 2023 — Google DeepMind's Gemini Ultra (Ultra 1.0), the largest and most capable model in the December 6, 2023 Gemini 1.0 family, built for highly complex tasks; Google reported state-of-the-art results on 30 of 32 widely used benchmarks, 90.0% on MMLU (the first model to beat human experts there) and 59.4% on MMMU. After further safety testing it reached consumers on February 8, 2024 through Gemini Advanced, part of the $19.99-a-month Google One AI Premium plan.
• Gemma Google DeepMind · LLM · 2024 — Google DeepMind's Gemma (February 21, 2024), Google's first family of lightweight open models, built from the research and technology behind Gemini and named after the Latin for 'precious stone'; it launched in 2B and 7B sizes, each pre-trained and instruction-tuned, under terms permitting responsible commercial use, and Google's report says it beat similarly sized open models on 11 of 18 text tasks. Gemma 1.1 followed on April 5, 2024, then CodeGemma and RecurrentGemma.
• Gemma 1 Google DeepMind · LLM · 2024 — Google DeepMind's Gemma (February 21, 2024), Google's first family of lightweight open models, built from the research and technology behind Gemini and named after the Latin for 'precious stone'; it launched in 2B and 7B sizes, each pre-trained and instruction-tuned, under terms permitting responsible commercial use, and Google's report says it beat similarly sized open models on 11 of 18 text tasks. Gemma 1.1 followed on April 5, 2024, then CodeGemma and RecurrentGemma.
• Gemma 1.0 2B Google DeepMind · LLM · 2024 — Google DeepMind's Gemma 2B, the smaller model of Google's first Gemma release: a lightweight open text-to-text decoder built from the research and technology behind Gemini, trained on 3 trillion tokens of primarily English web documents, mathematics and code with an 8,192-token context, and released in pre-trained and instruction-tuned versions; Google reports Gemma outperformed similarly sized open models on 11 of 18 text tasks.
• Gemma 1.0 7B Google DeepMind · LLM · 2024 — Google DeepMind's Gemma 7B, the larger model of Google's first Gemma release: a lightweight open text-to-text decoder built from the research and technology behind Gemini, trained on 6 trillion tokens of primarily English web documents, mathematics and code with an 8,192-token context, and released in pre-trained and instruction-tuned versions; Google reports Gemma outperformed similarly sized open models on 11 of 18 text tasks.
• Gemma 2 Google DeepMind · LLM (Open Weight) · 2024 — Google DeepMind's Gemma 2, the second generation of its lightweight open-weights models built from Gemini research: 9B and 27B text models released June 27, 2024 and a 2B added on July 31, on a redesigned architecture with alternating local and global attention, an 8,192-token context and knowledge distillation for the smaller sizes; Google said the 27B rivals models more than twice its size, and its report says Gemma 2 set a new state of the art for open models on the LMSYS Chatbot Arena.
• Gemma 2 27B Google DeepMind · LLM · 2024 — Google DeepMind's Gemma 2 27B, the largest model of the June 2024 Gemma 2 open-weights family: a text-only decoder trained without distillation on 13 trillion tokens of primarily English web, code and science data, with an 8,192-token context that alternates local sliding-window (4,096-token) and global attention; Google reports it ranked above Llama 3 70B in Chatbot Arena (Elo 1218 vs 1206) and runs at full precision on a single H100, A100 80GB or TPU host.
• Gemma 2 2B Google DeepMind · LLM · 2024 — Google DeepMind's Gemma 2 2B, the smallest Gemma 2 open-weights model, added July 31, 2024: a text-only decoder with about 2.6B parameters trained by knowledge distillation on 2 trillion tokens of primarily English data, with an 8,192-token context that alternates local sliding-window and global attention; Google reports it ranked above GPT-3.5-Turbo-0613 in Chatbot Arena (Elo 1126 vs 1116), and it comes in pre-trained and instruction-tuned versions.
• Gemma 2 9B Google DeepMind · LLM · 2024 — Google DeepMind's Gemma 2 9B, the mid-sized model of the June 2024 Gemma 2 open-weights family: a text-only decoder trained by knowledge distillation on 8 trillion tokens of primarily English web, code and science data, with an 8,192-token context that alternates local sliding-window and global attention; Google reports it outperforms Llama 3 8B and scored level with GPT-4-0314 in Chatbot Arena (Elo 1187 vs 1186), and the pre-trained model scores 71.3 on MMLU.
• Gemma 3 Google DeepMind · LLM · 2025 — Google DeepMind's Gemma 3 (March 12, 2025), open models in 1B, 4B, 12B and 27B sizes that run on a single GPU or TPU, with image input on the 4B, 12B and 27B, a 128K-token context (32K for the 1B), function calling and support for over 140 languages; distillation and a new post-training recipe made Gemma 3 4B-IT competitive with Gemma 2 27B-IT and Gemma 3 27B-IT comparable to Gemini 1.5 Pro, Google's report says. A 270M size followed in August 2025.
• Gemma 3 12B Google DeepMind · Multimodal LLM · 2025 — Google DeepMind's Gemma 3 12B, a multimodal open model from the March 2025 Gemma 3 family that takes text and images (896x896, encoded to 256 tokens by a 417M-parameter vision encoder) with a 128K-token context and support for over 140 languages; trained on 12 trillion tokens with distillation, it comes in pre-trained and instruction-tuned versions, and the instruction-tuned model scores 40.9 on GPQA Diamond and 85.7 on BIG-Bench Hard.
• Gemma 3 1B Google DeepMind · LLM · 2025 — Google DeepMind's Gemma 3 1B, the smallest model in the March 2025 Gemma 3 release and a text-only one with no vision encoder (302M embedding and 698M non-embedding parameters), with a 32K-token context and training on 2 trillion tokens with distillation; like the rest of Gemma 3 it interleaves five local sliding-window attention layers per global layer to cut KV-cache memory, and it comes in pre-trained and instruction-tuned versions.
• Gemma 3 270M Google DeepMind · LLM · 2025 — Google DeepMind's Gemma 3 270M (August 14, 2025), the smallest Gemma 3 model, built for task-specific fine-tuning, with 170M embedding parameters for its 256K-token vocabulary and 100M in its transformer blocks; trained on 6 trillion tokens with a 32K-token context, it comes in pre-trained, instruction-tuned and INT4 quantization-aware versions, and Google says the INT4 model used 0.75% of a Pixel 9 Pro's battery for 25 conversations, making it the most power-efficient Gemma.
• Gemma 3 27B Google DeepMind · Multimodal LLM · 2025 — Gemma 3 27B, the largest of Google DeepMind's Gemma 3 open models (March 12, 2025), built from the research behind Gemini 2.0 and trained on 14 trillion tokens: it takes text and image input with a 128K-token context, supports over 140 languages, function calling and structured output, and fits on a single GPU or TPU host. Google said it beat Llama3-405B, DeepSeek-V3 and o3-mini in early LMArena preference tests; weights, including official quantized versions, use the Gemma Terms of Use.
• Gemma 3 4B Google DeepMind · LLM · 2025 — Google DeepMind's Gemma 3 4B, a multimodal open model from the March 2025 Gemma 3 family that takes text and images (896x896, encoded to 256 tokens by a 417M-parameter vision encoder) with a 128K-token context and support for over 140 languages; trained on 4 trillion tokens with distillation and offered pre-trained and instruction-tuned, its instruction-tuned version is competitive with the much larger Gemma 2 27B IT, according to Google.
• Gemma 3n Google DeepMind · LLM · 2025 — Google DeepMind's Gemma 3n (June 2025), mobile-first open models for phones, laptops and tablets that take text, image, video and audio input; the MatFormer architecture and Per-Layer Embedding caching let the E2B and E4B models (5B and 8B raw parameters) run with the memory of 2B and 4B models, in as little as 2GB and 3GB. Gemma 3n has a 32K-token context and covers 140 languages for text, and E4B was the first model under 10B parameters to score over 1300 on LMArena.
• Gemma 4 Google DeepMind · Multimodal LLM · 2026 — Gemma 4 (April 2, 2026), Google DeepMind's most intelligent open model family, built from Gemini 3 research and released under Apache 2.0. It launched in four sizes: E2B and E4B edge models with 128K context and audio input that run offline on phones and Raspberry Pi, and a 26B mixture-of-experts (3.8B active) and 31B dense model with 256K context; a 12B came later. All read images and video, cover 140+ languages and support function calling; the 31B scores 89.2% on AIME 2026.
• Gemma 4 12B Google DeepMind · Multimodal LLM · 2026 — Encoder-free multimodal 12B targeting 16GB-VRAM local setups: one unified network instead of bolt-on vision/audio encoders. Stealth-updated Jul 15, 2026 (Flash Attention 4 on Hopper — 25-70% prefill speedup, tool-calling fixes, configurable vision resolution) with no version bump, drawing criticism about checkpoint ambiguity.
• Genie 2 Google DeepMind · World Model · 2024 — Google DeepMind's Genie 2 (December 2024), a research foundation world model that turns a single prompt image into a playable, action-controllable 3D environment that a person or AI agent can explore with keyboard and mouse for up to a minute; an autoregressive latent diffusion model trained on a large video dataset, it shows emergent object interactions, character animation, physics, lighting and memory of off-screen areas, and was built to train and evaluate embodied agents.
• Google DeepMind DeepTransformers Google DeepMind · LLM · 2017 — Attention Is All You Need.' Eight authors, one paper, and every AI model you have ever used descends from it. June 2017.
• Google DeepMind Gemini Robotics On-Device Google DeepMind · Multimodal LLM · 2025 — Gemini Robotics On-Device, Google DeepMind's vision-language-action model for bi-arm robots that runs entirely on the robot, without a data network, for latency-sensitive work and spotty connections. It builds on Gemini Robotics' generalization and dexterity, follows natural-language instructions and does tasks like unzipping bags and folding clothes, and was DeepMind's first VLA model offered for fine-tuning, adapting to new tasks from 50-100 demonstrations. Trusted testers get it with an SDK.
• Google DeepMind Gemini Robotics-ER 1.5 Google DeepMind · Multimodal LLM · 2025 — Thinking vision-language model for embodied reasoning that plans multi-step robot tasks, estimates progress and natively calls tools such as Google Search or a VLA model; the first Gemini Robotics model broadly available to developers, via the Gemini API.
• Google DeepMind Gemini Ultra Google DeepMind · Multimodal LLM · 2023 — Google DeepMind's Gemini Ultra (Ultra 1.0), the largest and most capable model in the December 6, 2023 Gemini 1.0 family, built for highly complex tasks; Google reported state-of-the-art results on 30 of 32 widely used benchmarks, 90.0% on MMLU (the first model to beat human experts there) and 59.4% on MMMU. After further safety testing it reached consumers on February 8, 2024 through Gemini Advanced, part of the $19.99-a-month Google One AI Premium plan.
• Imagen 3 Google DeepMind · Image Generation · 2024 — Google DeepMind's Imagen 3, a latent diffusion text-to-image model that its August 2024 technical report says was preferred over other state-of-the-art models at the time of evaluation; a December 2024 update made images brighter and better composed, widened its art styles from photorealism to impressionism, abstract and anime, and improved prompt following and detail. It powered ImageFX, Whisk and Vertex AI before Imagen 4 arrived in May 2025.
• Imagen 4 Google DeepMind · Image Generation · 2025 — Google DeepMind's Imagen 4 (May 20, 2025), a text-to-image model with sharper fine detail such as fabrics, water droplets and animal fur, in photorealistic and abstract styles, a range of aspect ratios at up to 2K resolution, and much better spelling and typography for cards and posters; it launched in the Gemini app, Whisk, Vertex AI and Google Workspace apps such as Slides, Vids and Docs, and Google announced a Fast variant up to 10x quicker than Imagen 3.
• Lyria 2 Google DeepMind · Music Generation · 2024 — Google DeepMind's Lyria 2, a music generation model for high-fidelity, professional-grade audio across genres and intricate compositions; it powered Music AI Sandbox, Google's experimental tools for musicians, producers and songwriters to create from text describing genres, moods, vocal styles and instruments, place lyrics on a timeline, and extend or edit tracks, then reached creators through YouTube Shorts and enterprises on Vertex AI, with SynthID watermarks on its output.
• Lyria 3 Google DeepMind · Music Generation · 2026 — Google DeepMind's Lyria 3 (February 18, 2026), a music generation model that writes lyrics for you, gives more control over style, vocals and tempo, and makes more realistic, musically complex tracks; it launched in beta in the Gemini app, turning text or images into 30-second songs in eight languages, with SynthID watermarks. Lyria 3 Pro, for songs of about three minutes with verses and choruses, and Lyria 3 Clip reached developers on March 25, 2026, and Lyria 3.5 followed in July.
• Lyria 3.5 Google DeepMind · Music Generation · 2026 — Full 3-minute cohesive songs with BPM and key-signature control in the prompt; more expressive multilingual vocals; style-transfer covers preserving a track's structure; lip-synced music videos via OmniFlash. All output SynthID-watermarked. No published benchmarks vs Suno/Udio; early testers say paid Suno 5.5 still edges it, but it's a free end-to-end create-to-publish stack.
• Med-Gemini Google DeepMind · Medical LLM · 2024 — Google's Med-Gemini, a family of multimodal Gemini models specialized for medicine that can use web search and be adapted to new modalities with custom encoders; Google reports state-of-the-art results on 10 of 14 medical benchmarks, 91.1% on MedQA (USMLE) and a 44.5% average relative gain over GPT-4V on 7 multimodal benchmarks, while saying more rigorous evaluation is needed before real-world use.
• MedGemma Google DeepMind · Science/Biology · 2025 — Google's MedGemma (May 20, 2025), open Gemma 3 variants trained for medical text and image comprehension: a 4B multimodal model and a 27B text-only model, with a 27B multimodal version added July 9; the multimodal models use a SigLIP encoder pre-trained on de-identified chest X-ray, dermatology, ophthalmology and histopathology images. MedGemma 27B text-only scores 87.7% on MedQA (64.4% for the 4B), and MedGemma 1.5 4B (January 2026) added CT, MRI and whole-slide pathology input.
• MusicLM Google DeepMind · Music Generation · 2023 — Google's MusicLM (January 2023), a research model that generates high-fidelity music from text descriptions by treating conditional music generation as a hierarchical sequence-to-sequence task; it produces 24 kHz audio that stays consistent over several minutes, beat earlier systems on audio quality and text adherence, can restyle whistled or hummed melodies to match a caption, and came with MusicCaps, 5.5K expert-captioned music clips.
• Nano Banana Google DeepMind · Image Generation · 2025 — Nano Banana, also known as Gemini 2.5 Flash Image, the image generation and editing model Google launched on August 26, 2025: it keeps a character or product consistent across prompts and edits, blends several images into one, makes targeted changes from plain-language instructions and draws on Gemini's world knowledge. It launched at $30 per million output tokens, about $0.039 per image, and every image carries an invisible SynthID watermark. The stable API model shuts down March 15, 2027.
• Nano Banana 2 Google DeepMind · Image Generation · 2026 — Launched as Google's flagship image generation and editing model, built on Gemini 3.1 Flash Image: 512px/1K/2K/4K output, native aspect ratios up to 8:1, web-search-grounded world knowledge, multi-language text rendering, and configurable 'thinking' levels. Google now recommends its update, Nano Banana 2.1.
• Nano Banana Pro Google DeepMind · Image Generation · 2025 — Google DeepMind's Nano Banana Pro (Gemini 3 Pro Image, November 20, 2025), an image generation and editing model built on Gemini 3 Pro that uses its reasoning, world knowledge and Google Search grounding for infographics and diagrams; it renders legible text in many languages, blends up to 14 images while keeping up to 5 people consistent, offers localized edits, lighting and camera controls, and outputs up to 4K. It launched in the Gemini app, Google Ads, Slides, Vids, AI Studio and Vertex AI.
• NanoBanana 2 Lite Google DeepMind · Image Generation · 2026 — Google DeepMind's Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image), its fastest and most cost-efficient Gemini image model, built for high-volume, near-real-time work: Google says it returns text-to-image results in 4 seconds at $0.034 per 1K-resolution image while keeping reliable prompt adherence, character consistency and legible text. It replaces the original Nano Banana for developers and also runs in AI Mode in Search, the Gemini app, NotebookLM, Google Photos and Flow.
• OmniFlash Google DeepMind · Video Generation · 2026 — First of Google's any-to-any Omni family: generates videos up to 10 seconds with precise conversational multi-turn editing ('make it daytime' redoes light, sky and shadows); editing Elo 1087. Also powers lip-synced music videos inside Flow Music.
• PaliGemma 2 Google DeepMind · Vision-Language Model · 2024 — Google's PaliGemma 2 (December 5, 2024), open vision-language models that pair the SigLIP-So400m image encoder with Gemma 2 language models of 2B, 9B and 27B, making 3B, 10B and 28B models trained at 224, 448 and 896-pixel resolutions; built for fine-tuning on captioning, visual question answering, OCR, detection and segmentation, they set state-of-the-art results on table, molecule and music-score recognition and radiography reports. Mix checkpoints followed in February 2025.
• RecurrentGemma Google DeepMind · LLM · 2024 — Google DeepMind's RecurrentGemma open-weights language models, built on the Griffin architecture, which replaces global attention with a mix of local attention and linear recurrences for fast inference on long sequences; released under the Gemma license with Flax (JAX) and PyTorch implementations.
• RT-2 Google DeepMind · Robotics · 2023 — Google DeepMind's RT-2, a vision-language-action model made by co-fine-tuning a web-scale vision-language model on robot trajectories, with robot actions written as text tokens; in 6,000 evaluation trials it generalized better to new objects, followed commands absent from its robot data and, with chain-of-thought, handled multi-stage reasoning such as choosing a rock as an improvised hammer.
• Sec-Gemini v1 Google DeepMind · Multimodal LLM · 2025 — Experimental cybersecurity model that combines Gemini with Google Threat Intelligence, OSV and other data sources; Google reports it beats other models by at least 11% on the CTI-MCQ threat-intelligence benchmark. Free to select research organizations.
• ShieldGemma Google DeepMind · Safety Model · 2024 — Google's ShieldGemma (July 31, 2024), open-weights safety classifiers built on Gemma 2 in 2B, 9B and 27B sizes that check user prompts and model responses against policies for four harms (sexually explicit content, dangerous content, hate and harassment) and answer Yes or No; trained mainly on synthetic data, the 9B beat Llama Guard by 10.8% in average AU-PRC on public benchmarks and topped WildGuard and GPT-4 in F1 by 4.3% and 6.4%.
• SigLIP Google DeepMind · Vision/CV · 2023 — Google's SigLIP, language-image pretraining with a simple pairwise sigmoid loss instead of CLIP's softmax contrastive loss, which works on image-text pairs alone without global normalization, scales to larger batches and also does better at small ones; a SigLiT model trained on four TPUv4 chips reached 84.5% ImageNet zero-shot accuracy in two days, and the models are released in Google's big_vision repository.
• T5Gemma Google DeepMind · LLM · 2025 — Google's T5Gemma (July 9, 2025), research encoder-decoder models made by adapting pretrained decoder-only Gemma 2 checkpoints (2B-2B, 9B-9B and unbalanced pairs such as 9B-2B) plus new T5-sized Small to XL models trained with the Gemma 2 recipe; Google reports they match or beat their decoder-only counterparts at similar inference cost, with the instruction-tuned 2B-2B nearly 12 points above Gemma 2 2B on MMLU. T5Gemma 2 followed in December 2025 in 270M-270M, 1B-1B and 4B-4B sizes.
• TranslateGemma Google DeepMind · Translation LLM · 2026 — Google's TranslateGemma (January 15, 2026), open translation models in 4B, 12B and 27B sizes, fine-tuned from Gemma 3 with supervised training on synthetic and human-translated parallel data and then reinforcement learning guided by MetricX-QE and AutoMQM reward models; covering 55 languages, the 12B beats the Gemma 3 27B baseline on WMT24++ by MetricX and the 4B rivals the 12B baseline. The models keep Gemma 3's image input for translating text in images and take a 2K-token context.
• TxGemma Google DeepMind · Science/Biology · 2025 — Google's TxGemma, open models fine-tuned from Gemma 2 on 7 million examples curated from the Therapeutics Data Commons to predict properties of small molecules, proteins, nucleic acids, diseases and cell lines across drug development; 2B, 9B and 27B 'predict' models are joined by 9B and 27B 'chat' models that can explain their reasoning. Across 66 therapeutic tasks, the 27B predict model matched or beat Google's earlier Tx-LLM on 64 and specialist models on 50.
• VaultGemma Google DeepMind · LLM · 2025 — Google's VaultGemma (September 2025), a 1B-parameter open model with a Gemma 2-like architecture that Google calls the largest open model trained from scratch with differential privacy, using DP-SGD with a sequence-level guarantee (ε ≤ 2.0) over 1,024-token sequences, guided by new DP scaling laws; tests found no detectable memorization. It ships as a pre-trained model with a 1,024-token context, and Google says its utility matches non-private models from about five years ago.
• Veo 2 Google DeepMind · Video Generation · 2024 — Google DeepMind's Veo 2 (December 16, 2024), a video generation model that beat leading models in human raters' head-to-head comparisons at release, with a better grasp of physics and human movement, an understanding of cinematography such as lenses and shot types, fewer hallucinated details like extra fingers, and resolutions up to 4K at lengths extending to minutes; it rolled out through VideoFX, YouTube and Vertex AI with SynthID watermarks, and Veo 3 succeeded it in May 2025.
• Veo 3 Google DeepMind · Video Generation · 2025 — Google DeepMind's Veo 3 (May 20, 2025), the first Veo model to generate video with sound, from background noise and birdsong to dialogue between characters, improving on Veo 2's quality with strong text and image prompting, real-world physics and accurate lip syncing; it launched for Google AI Ultra subscribers in the U.S. in the Gemini app and Flow and for enterprises on Vertex AI. Veo 3.1 succeeded it in October 2025.
• Veo 3.1 Google DeepMind · Video Generation · 2025 — Google DeepMind's Veo 3.1 (October 15, 2025), an update to Veo 3 with richer generated audio, more narrative control, more realistic textures, stronger prompt adherence and better audiovisual quality when turning images into video; in Google's Flow filmmaking tool it brought sound to Ingredients to Video (multiple reference images), Frames to Video (first and last frames) and Extend (shots of a minute or more). It is also in the Gemini API, Vertex AI and the Gemini app.
• Transformer-XL Google/CMU · LLM · 2019 — Transformer-XL, a language-model architecture that adds segment-level recurrence and a new positional encoding scheme so Transformers can learn dependencies beyond a fixed-length context; it learns dependencies 80% longer than RNNs and 450% longer than vanilla Transformers, evaluates up to 1,800+ times faster, and set state-of-the-art results on enwik8, text8, WikiText-103, One Billion Word and Penn Treebank.
• Grammarly AI Grammarly · AI Writing · 2023 — AI writing assistant. Grammar, style, tone.
• Granola Granola · Meeting AI · 2024 — AI notepad for meetings.
• Groq LPU Inference Groq · Inference Platform · 2024 — Ultra-fast LLM inference on custom chips.
• Holo 3.1 H Company · Agent · 2026 — Family of local computer-use agent models from 0.8B to 35B parameters, targeting screen-driving agents on local hardware rather than in the cloud.
• Haiper Haiper · Video Generation · 2024 — 2-second video clips from text/image.
• Harvey AI Harvey · Legal AI · 2023 — AI for legal professionals.
• Hex AI Hex · Data Analysis · 2023 — AI-native data workspace.
• HeyGen HeyGen · Video Generation · 2023 — AI avatar videos. Translation and dubbing.
• Huawei Dream 7B Huawei · LLM · 2025 — Open 7B diffusion language model that generates text by iterative denoising, developed by the University of Hong Kong with Huawei Noah's Ark Lab. Apache 2.0.
• Huawei Pangu Ultra Huawei · LLM · 2025 — 135B-parameter dense LLM trained on Huawei's Ascend NPUs over 13.2T tokens, using depth-scaled sandwich normalization to eliminate loss spikes, with reasoning strengthened in post-training.
• PanGu-Σ Huawei · LLM · 2023 — 1.085-trillion-parameter sparse LLM trained on Ascend 910 processors with MindSpore, extending PanGu-α into a Random Routed Experts model trained on 329B tokens.
• PixArt-α Huawei · Image Generation · 2023 — PixArt-α, a 0.6B-parameter diffusion transformer for text-to-image generation that produces 1024px images using a T5 text encoder and a VAE; its authors report it took 10.8% of Stable Diffusion v1.5's training time (675 versus 6,250 A100 GPU days) and that users preferred it over SDXL 0.9, Stable Diffusion 2, DALL-E 2 and DeepFloyd.
• PixArt-Σ Huawei · Image Generation · 2024 — PixArt-Σ, a 0.6B-parameter diffusion transformer for text-to-image generation, trained weak-to-strong from PixArt-α, that generates 1024px, 2K and 4K images directly in a single sampling process using a T5 text encoder and a VAE.
• HubSpot AI HubSpot · Marketing AI · 2023 — AI tools for marketing and sales.
• BLOOM Hugging Face · LLM · 2022 — 176B parameters. 46 languages. Built by 1,000+ researchers across the world. BigScience collaboration. The people's model. July 2022.
• BLOOM-176B Hugging Face · LLM · 2022 — 176B parameters. Full BLOOM model. Open-source multilingual foundation model.
• BLOOMZ Hugging Face · LLM · 2022 — BigScience's 176B BLOOMZ, the BLOOM multilingual model fine-tuned on the xP3 crosslingual task mixture so it follows instructions zero-shot in dozens of languages; the family also includes smaller BLOOMZ and mT0 models, with xP3mt-tuned versions recommended for non-English prompts.
• DistilBERT Hugging Face · LLM · 2019 — Hugging Face's DistilBERT, a distilled version of BERT base (uncased) pretrained on BERT's corpus with BERT base as the teacher; its paper reports it is 40% smaller and 60% faster while keeping 97% of BERT's language-understanding ability.
• Hugging Face Hub Chat Hugging Face · Chat Interface · 2023 — Free access to open models on HF.
• Idefics2 Hugging Face · Multimodal LLM · 2024 — Hugging Face's Idefics2, an 8B open multimodal model (SigLIP vision encoder plus Mistral-7B) that takes interleaved images and text and returns text; it improves on Idefics1 in OCR, document understanding and visual reasoning and comes in base, instruction-tuned and chatty versions.
• Idefics3 Hugging Face · Multimodal LLM · 2024 — Hugging Face's Idefics3, an 8B open multimodal model (SigLIP vision encoder plus Llama 3.1 8B Instruct) that takes interleaved images and text and returns text; it improves on Idefics1 and Idefics2 in OCR, document understanding and visual reasoning and was post-trained with supervised fine-tuning only, without RLHF.
• Parler-TTS Hugging Face · Speech Synthesis · 2024 — Hugging Face's Parler-TTS, a lightweight, fully open-source text-to-speech model that speaks in a style described in text (gender, pitch, speaking style and more), reproducing research by Stability AI and the University of Edinburgh; the August 2024 Mini (880M parameters) and Large (2.3B) versions were trained on 45K hours of audiobook data, with datasets, training code and weights all released.
• SmolLM Hugging Face · LLM · 2024 — Hugging Face's SmolLM small language models (135M, 360M and 1.7B parameters), trained on the Cosmo-Corpus of synthetic textbooks and stories (Cosmopedia v2), educational Python (Python-Edu) and educational web pages (FineWeb-Edu); they work mainly in English.
• SmolLM2 Hugging Face · LLM · 2024 — Hugging Face's SmolLM2 compact language models (135M, 360M and 1.7B parameters), light enough to run on-device; the 1.7B model was trained on 11 trillion tokens and improves on SmolLM in instruction following, knowledge, reasoning and math, and its instruct version supports rewriting, summarization and function calling.
• SmolVLM Hugging Face · Multimodal LLM · 2024 — Hugging Face's SmolVLM, a compact open multimodal model built on SmolLM2 and the Idefics3 architecture that takes interleaved images and text and returns text; it encodes each 384x384 image patch as 81 visual tokens for faster, lower-memory on-device use.
• Zephyr 7B Hugging Face · LLM · 2023 — Hugging Face's Zephyr-7B-β chat model: Mistral-7B-v0.1 fine-tuned with Direct Preference Optimization on public synthetic datasets; at release it was the highest-ranked 7B chat model on MT-Bench (7.34) and AlpacaEval (90.60% win rate). Removing the datasets' built-in alignment raised its scores but makes problematic output more likely.
• Hume EVI 2 Hume AI · Emotionally Intelligent Voice · 2024 — Hume's second-generation Empathic Voice Interface, a real-time speech-to-speech voice AI that measured the user's vocal expression and responded with empathic speech. Hume deprecated EVI 1 and 2 on August 30, 2025 in favor of EVI 3, and is shutting down its whole EVI and TTS APIs on November 13, 2026.
• Bamba-9B-v2 IBM · LLM · 2025 — Mamba-2-based 9B decoder-only model trained on an extra 1T tokens over Bamba v1; IBM says it beats Llama 3.1 8B on L1/L2 leaderboard scores despite nearly 5x less training data. Apache 2.0.
• Granite 13b IBM · LLM · 2023 — IBM's 13-billion-parameter Granite language model for watsonx.ai, trained on enterprise-relevant data from internet, academic, code, legal and finance sources; offered as granite-13b-chat-v2 for dialog and granite-13b-instruct-v2.
• Granite 3.1 8B IBM · LLM · 2025 — 8B long-context Granite instruct model fine-tuned from Granite-3.1-8B-Base on permissively licensed open and synthetic data. Apache 2.0.
• Granite 3.3 8B Instruct IBM · LLM · 2025 — 8B, 128K-context Granite model fine-tuned for better reasoning and instruction following, with structured reasoning through dedicated think/response tags. Apache 2.0.
• Granite 4.0 IBM · LLM · 2025 — IBM's hybrid Mamba-2/Transformer family: Granite-4.0-H-Small (32B total, ~9B active MoE), H-Tiny (7B, ~1B active), H-Micro (3B hybrid) and Micro (3B transformer), all Apache 2.0. IBM cites >70% lower GPU memory use, ISO 42001 certification and cryptographically signed checkpoints; trained on 22T tokens.
• Granite Code Models IBM · Code LLM · 2024 — IBM's Granite Code family of open code models in 3B, 8B, 20B and 34B sizes, each with base and instruct versions; the instruct models are fine-tuned on permissively licensed instruction data such as Git commits paired with human instructions. Apache 2.0.
• Granite-3.2-8B-Instruct IBM · LLM · 2025 — 8B long-context Granite model fine-tuned for reasoning ('thinking') that can be switched on only when needed (Feb 2025). Apache 2.0.
• Granite-4.0-Tiny-Preview IBM · LLM · 2025 — Preview of IBM Granite 4.0: a 7B-parameter fine-grained hybrid mixture-of-experts instruct model, tuned with permissively licensed open and synthetic data for long-context tasks. Apache 2.0.
• ICONNAI ICONN-1 ICONNAI · LLM · 2025
• Grounding DINO IDEA Research · Vision/CV · 2023 — IDEA Research's Grounding DINO, an open-set object detector that pairs the DINO detector with grounded pre-training so it can detect arbitrary objects described in language; it scored 52.5 AP zero-shot on COCO without COCO training data and 63.0 AP after fine-tuning. IDEA has since released Grounding DINO 1.5 through an API.
• Ideogram 1.0 Ideogram · Image Generation · 2023 — Text-to-image with strong typography.
• Ideogram 2.0 Ideogram · Image Generation · 2024 — Ideogram's August 2024 model, pitched at launch for its text rendering in images (typography, logos, signs); Ideogram's site now lists only Ideogram 4.5, 4.0 and 3.0 among its models.
• Ideogram 3.0 Ideogram · Image Generation · 2025 — Ideogram's 3.0 image model, focused on photorealistic and design images; since succeeded by Ideogram 4.0.
• Ideogram 4.0 Ideogram · Image Generation · 2026 — 9.3B-parameter text-to-image with open weights under a non-commercial license; leads open-weight image models on typography and layout, with bounding-box/layout-style prompting that trades casual ease for precise structured control.
• SparkDesk iFlytek · Chat Interface · 2023 — iFlytek's AI assistant. Chinese market.
• K2 Horizon 0.9B IFM · LLM · 2026 — Watch-scale 0.9B dense model - best at its size for math and reasoning, 131,072-token context, Apache 2.0. Runs on wearables.
• K2 Horizon 3.7B IFM · LLM · 2026 — Fine-tuning-friendly 3.7B dense model with industry-leading reasoning under 4B parameters, 524K context, Apache 2.0.
• K2 Horizon 32B IFM · LLM · 2026 — Dense 32B K2 Horizon model for local hosting and on-prem servers, 524K context, Apache 2.0. Fully open.
• K2 Horizon 36B-A4B IFM · LLM · 2026 — Sparse Mixture-of-Value-Attention model in the K2 Horizon fleet: 36B total / 4B active, 524K context, Apache 2.0. Fully open (weights, code, training data).
• K2 Horizon 375B-A23B IFM · LLM · 2026 — Flagship of IFM's fully-open K2 Horizon fleet: 375B MoE / 23B active, 524,288-token context, Apache 2.0 - released with full weights, code and training data. Targets enterprise reasoning and demanding workloads.
• K2 Horizon 7B IFM · LLM · 2026 — Best-performing open model under 10B at release: 7B dense, 524K context, phone-compatible, Apache 2.0. Fully open.
• Inception Mercury Inception · LLM · 2025 — Diffusion LLM (dLLM): generates text in parallel rather than token by token, which Inception says makes it up to 10x faster and more efficient than conventional LLMs. Part of the Feb 2025 launch of the first commercial-scale dLLMs.
• Inception Mercury Coder Small Inception · LLM · 2025 — Code-focused Mercury diffusion LLM (dLLM) from the Feb 2025 launch of the first commercial-scale dLLMs; generates code in parallel rather than token by token for much faster responses.
• InclusionAI Ring-1T-preview InclusionAI · LLM · 2025 — Preview (late Sep 2025) of InclusionAI's trillion-parameter thinking model Ring-1T; the full open-source Ring-1T followed in Oct 2025 under the MIT license.
• Inflection Pi Inflection AI · Chat Interface · 2023 — Inflection Pi, the Pi personal AI from Inflection AI (founded in 2022 by Mustafa Suleyman, Karén Simonyan and Reid Hoffman), first released May 2, 2023 as a kind, supportive companion that acts as a coach, confidant, creative partner and sounding board, free at launch on the web, iPhone, WhatsApp, SMS, Instagram and Messenger. Pi now runs on Inflection's custom fine-tuned models with voice, memory and reminders, and Inflection AI Labs added the Pi Journeys experiment in July 2026.
• Inflection-2.5 Inflection AI · LLM · 2024 — Inflection-2.5 (March 2024), Inflection AI's upgraded model behind its Pi personal AI: Inflection said it approached GPT-4's performance, reaching more than 94% of GPT-4's average performance while using only 40% of the training compute, and reported results on MMLU, GPQA, BIG-Bench-Hard, HumanEval+ and other benchmarks. It launched with real-time web search in Pi, then used by 1 million people daily and 6 million monthly, and Inflection pairs its IQ with emotional intelligence (EQ).
• Inflection-3.0 Inflection AI · LLM · 2024 — Inflection 3.0 (October 2024), the enterprise-grade AI system behind Inflection for Enterprise, which lets a company own its data, its fine-tuned model and the architecture it runs on, hosted on-premises, in the cloud or hybrid. It runs on Intel Gaudi 3 accelerators and Intel Tiber AI Cloud, which Inflection says gives up to 2x better price performance; Inflection launched a commercial API with it and later that month added UiPath-powered Agentic Workflows.
• Pi Inflection AI · Chat Interface · 2023 — Pi ('personal intelligence'), Inflection AI's personal AI, launched in May 2023 as a kind and supportive companion for conversation, advice and concise answers. Inflection now calls it the world's first personal intelligence: it chats by text or voice, remembers lasting details across chats, sets reminders and to-dos, offers a personalized Daily Brief of news and runs on Inflection's custom fine-tuned models. It is free on the web, iOS and Android, and cannot generate, accept or edit images.
• Neural Chat Intel · LLM · 2023 — Intel's NeuralChat, a series of 7B chat models that Intel's NeuralChat team fine-tuned on Intel Gaudi 2 accelerators from Mistral-7B-v0.1, using the SlimOrca instruction dataset and Direct Preference Optimization on Intel's orca_dpo_pairs; v3 and v3-1 came in October 2023 and v3-3, further tuned on the MetaMathQA math dataset, in December 2023. 8,192-token context; Apache 2.0.
• Neural Chat 7B Intel · LLM · 2023 — Intel's Neural-Chat 7B v3-1 (October 2023), a 7B-parameter chat model that Intel's NeuralChat team fine-tuned on Intel Gaudi 2 processors from Mistral-7B-v0.1 on the open SlimOrca dataset and aligned with Direct Preference Optimization on Intel's orca_dpo_pairs, keeping Mistral's 8,192-token context; Intel later released v3-3 with added math tuning on MetaMathQA. Apache 2.0.
• Intercom Fin Intercom · Support AI · 2023 — AI customer support chatbot.
• Moxin-7B International Consortium · LLM · 2024 — Moxin-7B, a family of fully open-source, reproducible 7B language models (Base, Instruct, Reasoning and VLM) released under the Model Openness Framework with their training data, training scripts and weights; the Instruct model was post-trained with Ai2's Tülu 3 recipe and the Reasoning model with reinforcement learning via DeepScaleR.
• International Moxin-7B International Research Consortium · LLM · 2025 — Moxin-7B from the Moxin project, a family of fully open-source, reproducible 7B language models (Base, Instruct, Reasoning and VLM) built to the Model Openness Framework, with the training datasets, training scripts and trained weights all released; post-training used the Tülu 3 recipe and GRPO reinforcement learning. Apache 2.0.
• Invideo AI Invideo · Video Generation · 2023 — Text-to-video with AI editing.
• InvokeAI InvokeAI · Image Generation · 2022 — Invoke (InvokeAI), a creative engine for generating, refining and iterating on images with AI through a web-based interface and workflows, installed locally with a launcher; free under a commercially friendly license and used as the foundation of several commercial products.
• Jan Jan AI · Local Inference · 2024 — Open-source ChatGPT alternative. Local.
• Jasper AI Jasper · AI Writing · 2023 — AI marketing content creation.
• Mellum 2 JetBrains · Code LLM · 2026 — 12B mixture-of-experts coding model with 2.5B active parameters, trained from scratch by a small team using a three-stage curriculum over 10T tokens — an IDE company converting years of developer-workflow context into model advantage.
• mmBERT JHU · LLM · 2025 — Multilingual ModernBERT-style encoder from JHU. 1,800+ languages, 3T tokens. First major multilingual encoder refresh since XLM-R.
• Jina Embeddings v3 Jina AI · Embeddings · 2024 — Jina AI's jina-embeddings-v3 (September 2024), a 570M-parameter multilingual embedding model built on Jina-XLM-RoBERTa with an 8,192-token context and five task-specific LoRA adapters (retrieval, clustering, classification and text matching); Matryoshka training lets its 1,024-dimension vectors be cut to as few as 32, it covers 89 languages (strongest in 30), and Jina reports it beats OpenAI's and Cohere's latest proprietary embeddings on English MTEB tasks. Weights CC BY-NC 4.0.
• Jina Reranker v2 Jina AI · Reranking · 2024 — Jina AI's Jina Reranker v2 (jina-reranker-v2-base-multilingual), a 278M-parameter multilingual cross-encoder that scores how relevant each document is to a query for search and retrieval-augmented generation, reading 1,024 tokens at a time and longer texts with a sliding window, with flash attention and added strength in function-calling-aware, text-to-SQL-aware and code retrieval reranking. Weights for research and evaluation under CC BY-NC 4.0; commercial use through Jina's API, AWS or Azure.
• RNN LM Johns Hopkins · LLM · 1986 — Recurrent neural network language model (RNNLM): Mikolov, Karafiát, Burget and Černocký (Brno University of Technology) with Khudanpur (Johns Hopkins) showed at Interspeech 2010 that recurrent-network LMs, which carry context in a hidden state, beat n-gram models: a mixture of them cut perplexity by about 50% versus a state-of-the-art backoff model, and word error rate fell about 18% on the Wall Street Journal task.
• Julius AI Julius · Data Analysis · 2023 — AI data analysis and visualization.
• Kagi AI Kagi · Search AI · 2023 — AI features in Kagi search.
• Kaiber Kaiber · Video Generation · 2023 — AI video creation and animation.
• Remove.bg Kaleido AI · Image Editing · 2018 — AI background removal. The standalone remove.bg site closes on December 1, 2026; background removal moves into Canva and the API to Leonardo.ai.
• Kapwing AI Kapwing · Video Editing · 2023 — AI video editing. Auto-captions, resize.
• Khan Academy Khanmigo Khan Academy · Education AI · 2023 — AI tutor powered by GPT-4.
• Zing (Krea AI) Krea AI · Image Generation · 2024 — Real-time AI image generation.
• Krisp Krisp · Audio Processing · 2023 — AI noise cancellation for calls.
• Kling 1.0 Kuaishou · Video Generation · 2024 — Kuaishou video model. 2-min clips.
• Kling 1.5 Kuaishou · Video Generation · 2024 — Improved video quality and motion.
• Kling 2.0 Kuaishou · Video Generation · 2025 — 2-minute videos. Camera controls. Character consistency. Bargain pricing.
• Kling 2.5 Turbo Kuaishou · Video Generation · 2025 — Text-to-video and image-to-video upgrade with better motion dynamics, prompt adherence and cinematic quality at lower cost. In Kuaishou's professional blind tests it beat Seedance 1.0 mini, Veo3-fast and Seedance 1.0 on text-to-video (win-loss ratios 285%, 212% and 160%).
• Kling 3.0 Kuaishou · Video Generation · 2026 — Native 4K at 60fps, clips up to 15 seconds, multilingual lip-sync and integrated audio in a single pipeline; storyboard mode for per-shot camera and pacing control; Omni One engine praised for material/fluid physics. Four entries in the AA top 10. Released Feb 4-5, 2026 (sources vary by timezone).
• Kuaishou KAT-V1-200B Kuaishou · LLM · 2025 — 200B-parameter mixture-of-experts (40B active) member of Kuaishou's Kwaipilot-AutoThink family, which learns when to reason step by step and when to answer directly; reported as still in training in the July 2025 KAT-V1 report.
• Kuaishou KAT-V1-40B Kuaishou · LLM · 2025 — Open-source 40B model (Kwaipilot-AutoThink) that learns to switch between step-by-step reasoning and direct answers based on task complexity, to curb over-thinking.
• Kuaishou Klear-46B-A2.5B Kuaishou · LLM · 2025 — Sparse mixture-of-experts LLM with 46B total and 2.5B active parameters (256 experts, with 8 plus a shared expert active per layer), from Kuaishou's Kwai-Klear team. Apache 2.0.
• Helium-1 Kyutai · Audio/Speech · 2025 — Lightweight 2B-parameter language model from Kyutai for edge and mobile devices, supporting the 24 official EU languages; a base model (not instruction-tuned), released under CC BY-SA 4.0.
• Moshi Kyutai · Speech AI · 2024 — Kyutai's Moshi, a speech-text foundation model and full-duplex spoken dialogue framework for real-time conversation, built on Mimi, Kyutai's streaming neural audio codec; it ships PyTorch (research), MLX (on-device on iPhone and Mac) and Rust (production) inference stacks.
• OpenAssistant LAION · LLM · 2023 — LAION-led open-source chat assistant project that crowd-sourced OpenAssistant Conversations, a human-written, human-rated assistant dataset released to democratize alignment research.
• Lalal.ai Lalal.ai · Audio Processing · 2023 — AI vocal/stem separation.
• Leonardo AI Leonardo AI · Image Generation · 2023 — AI art platform with fine-tuned models.
• EXAONE 4.0 LG · LLM · 2025 — LG AI Research's EXAONE 4.0 combines a non-reasoning mode and a reasoning mode in one model, adds agentic tool use and supports English, Korean and Spanish; it comes as a 32B model (hybrid local and global attention, 131K context) and a 1.2B on-device model.
• EXAONE 4.5 LG · Multimodal LLM · 2026 — LG AI Research's open-weight vision-language update in the EXAONE 4 line, aimed at industrial and enterprise intelligence. Multimodal (text + vision).
• EXAONE Deep LG · LLM · 2025 — LG AI Research's reasoning-focused EXAONE Deep models (2.4B, 7.8B and 32B) for math and coding; the 32B model has 30.95B non-embedding parameters, 64 layers and a 32K context. LG reports the 7.8B model outperforms OpenAI's o1-mini.
• EXAONE-3.5 LG · LLM · 2025 — LG AI Research's bilingual (English and Korean) instruction-tuned EXAONE 3.5 models in 2.4B, 7.8B and 32B sizes, all with a 32K context; the 32B model has 30.95B non-embedding parameters.
• LTX-Video Lightricks · Video Generation · 2024 — First DiT-based open-weights text-to-video model. 2B params at launch (Nov 2024); LTXV-13B added May 2025; real-time generation on consumer GPUs. Runs locally via ComfyUI. Succeeded by LTX-2 (open audio+video, Jan 2026).
• LinkedIn EON-8B LinkedIn (Microsoft) · LLM · 2025 — LinkedIn's in-house model from its Economic Opportunity Network (EON) project: Llama 3.1 8B domain-adapted with LinkedIn data. LinkedIn says it is 75x cheaper to serve than GPT-4 and 6x cheaper than GPT-4o, and it uses EON models in features such as Hiring Assistant.
• LFM2.5-2.6B Liquid AI · LLM · 2026 — 2.69B-parameter hybrid (128K context) pre-trained on ~34T tokens; post-training ran agentic RL inside real harnesses (Hermes Agent, OpenClaw, Pi) so tool calling was learned where it happens. Beats Qwen3.5-9B — three times its size — on ToolSandbox (77.83) and instruction following; 220 tok/s on an M5 Max CPU. Model card honestly scopes it away from agentic coding and knowledge-heavy work: built for private, on-device agents.
• Amber LLM360 · LLM · 2023 — LLM360's Amber, a 7B English language model with the LLaMA architecture, released with 360 training checkpoints and its full data sequence under Apache 2.0; LLM360 says it is not a state-of-the-art model but is meant to make LLM training knowledge accessible.
• LM Studio LM Studio · Local Inference · 2023 — Desktop app for running local LLMs.
• Vicuna LMSYS · LLM · 2023 — LMSYS's Vicuna, an open chatbot introduced March 30, 2023 that fine-tuned Meta's LLaMA on about 70,000 user-shared ChatGPT conversations from ShareGPT with a 2,048-token context; the 13B model trained in a day on eight A100 GPUs for about $300 and, in a preliminary test judged by GPT-4, reached more than 90% of ChatGPT and Bard quality. Weights were released for non-commercial use, and the August 2023 v1.5 moved to Llama 2 with 4K and 16K contexts.
• Vicuna-13B LMSYS · LLM · 2023 — LMSYS's Vicuna v1.5 13B chat assistant, fine-tuned from Llama 2 with supervised instruction tuning on about 125K user-shared ShareGPT conversations; intended mainly for research on LLMs and chatbots.
• Loom AI Loom/Atlassian · Video Communication · 2023 — AI video messaging with summaries.
• Loudly AI Loudly · Music Generation · 2023 — AI music generator. Royalty-free.
• Lovable.dev Lovable · Code AI · 2024 — Lovable, which describes itself as an AI software engineer that lets anyone build for the web: users chat to build websites and web apps with no technical knowledge needed, and credits pay for building, hosting apps on Lovable Cloud and AI features inside those apps. The free plan includes 5 build credits a day (up to 30 a month) plus 20 Cloud credits a month; paid Pro and Business plans add a monthly credit balance, and top-up credits last 12 months.
• Replika Luka · Chat Interface · 2017 — AI companion chatbot.
• Luma Dream Machine Luma AI · Video Generation · 2024 — Text/image-to-video. Natural language editing. Reframe feature.
• Luma Genie Luma AI · Video Generation · 2024 — Dream Machine. Text to video.
• Luma Ray2 Luma AI · Video Generation · 2025 — Next-gen video model. Improved coherence.
• Luma Ray3 Luma AI · Video Generation · 2025 — First 'reasoning' video model with native 16-bit HDR generation; Ray3 Modify for video-to-video editing of actor footage. Ray3.14 update Jan 26, 2026 (with draft-mode cost cut); Ray3.2 (2026) extends to 20s/1080p with 16-bit EXR export for grading/compositing pipelines.
• Magnific AI Magnific · Image Enhancement · 2023 — AI image upscaling and enhancement.
• Mailchimp AI Mailchimp · Marketing AI · 2023 — AI email marketing optimization.
• Make.com AI Make · Automation AI · 2023 — AI-enhanced workflow automation.
• MBZUAI K2-Think MBZUAI · LLM · 2025 — 32B reasoning system built on Qwen2.5 that MBZUAI reports matches or beats much larger models such as GPT-OSS 120B and DeepSeek V3.1, combining long chain-of-thought fine-tuning, RL with verifiable rewards, agentic planning and test-time scaling.
• Jais 30B MBZUAI/Inception · LLM · 2023 — 30B Arabic-English bilingual model.
• LongCat-2.0 Meituan · LLM · 2026 — 1.6-trillion-parameter MoE trained entirely on Chinese ASICs without NVIDIA hardware: 59.5 SWE-bench Pro. Served anonymously as 'Owl Alpha' before disclosure and ranks among OpenRouter's top models by volume — part of the surge putting Chinese open-weight models at ~30% of global usage, up from 1.2% eleven months earlier.
• LongCat-Flash-Lite-Sparse Meituan · LLM · 2026 — 69B total / ~3B active per token with native 1M-token context (up from 256K dense), MIT licensed. LongCat Sparse Attention lifts SWE-Bench Verified from 54.4 to 68.2 and SWE-Bench Multilingual by 21 points over the dense twin, building on DeepSeek Sparse Attention while removing its O(L^2) scoring overhead; paper reports the architecture scaling to 560B-A27B.
• Mem AI Mem · Productivity AI · 2023 — AI-powered note-taking.
• Meshy Meshy · 3D Generation · 2023 — Text/image to 3D meshes. Game-ready.
• Galactica Meta · LLM · 2022 — Meta AI's Galactica (November 2022), decoder-only language models for science from 125M to 120B parameters, built by the Papers with Code team and trained on 106 billion tokens of open-access scientific text and data such as papers, textbooks, encyclopedias and knowledge bases; it beat GPT-3 on LaTeX equation probes (68.2% vs 49.0%) and PaLM 540B on MATH (20.4% vs 8.8%). Released under CC BY-NC 4.0, with Meta cautioning against production use without safeguards because it can hallucinate.
• OPT Meta · LLM · 2022 — Meta AI's OPT (Open Pre-trained Transformers, May 3, 2022), decoder-only language models from 125M to 175B parameters trained on publicly available data; OPT-175B is comparable to GPT-3 while needing 1/7th the carbon footprint to develop. Meta shared OPT-175B with researchers under a noncommercial license, calling it the first release at that scale to include both the trained model and the code to train and use it, fully released the smaller models, and published its training logbook.
• AudioCraft Meta AI · Audio Generation · 2023 — Meta's AudioCraft, a PyTorch library for audio-generation research with training and inference code for MusicGen (text to music), AudioGen (text to sound), the EnCodec neural audio codec, Multi Band Diffusion, MAGNeT, AudioSeal watermarking, MusicGen Style and JASCO.
• BLT Meta AI · LLM · 2024 — Byte-level LLM architecture that groups raw bytes into dynamically sized patches based on next-byte entropy, matching tokenization-based models at scale with better inference efficiency and robustness; studied up to 8B parameters.
• Chameleon Meta AI · Multimodal LLM · 2024 — Meta FAIR's Chameleon (May 2024 paper), early-fusion, token-based models that understand and generate interleaved images and text in any order with one architecture; Meta reports state-of-the-art image captioning, better text-only results than Llama 2, parity with Mixtral 8x7B and Gemini Pro, and human ratings matching or beating Gemini Pro and GPT-4V on long mixed-modal output. In June 2024 Meta released safety-tuned 7B and 34B models under a research-only license, with text-only output.
• CM3Leon Meta AI · Multimodal · 2023 — Meta's CM3Leon, a retrieval-augmented, token-based, decoder-only multimodal model that generates and infills both text and images, trained with a recipe adapted from text-only LLMs (retrieval-augmented pretraining, then multi-task supervised fine-tuning); Meta reports state-of-the-art text-to-image results (zero-shot MS-COCO FID of 4.88) with 5x less training compute than comparable methods, plus language-guided image editing.
• Code Llama Meta AI · Code LLM · 2023 — Meta's code-specialized Llama family in 7B, 13B, 34B and 70B sizes, each in base, Python and Instruct variants. The 7B, 13B and 70B models support code infilling, and all but the 70B Python and Instruct models were fine-tuned on up to 16K tokens and handle up to 100K tokens at inference.
• Code Llama 13B Meta AI · Code LLM · 2023 — 13B model in Meta's Code Llama family, in base, Python and Instruct variants; supports code infilling, fine-tuned on up to 16K tokens and handles up to 100K tokens at inference.
• Code Llama 34B Meta AI · Code LLM · 2023 — 34B model in Meta's Code Llama family, in base, Python and Instruct variants; fine-tuned on up to 16K tokens and handles up to 100K tokens at inference.
• Code Llama 70B Meta AI · Code LLM · 2024 — 70B model in Meta's Code Llama family, in base, Python and Instruct variants. The base model supports code infilling and, like the smaller models, was fine-tuned on up to 16K tokens and handles up to 100K tokens at inference; the 70B Python and Instruct variants were not.
• Code Llama 7B Meta AI · Code LLM · 2023 — 7B model in Meta's Code Llama family, in base, Python and Instruct variants; supports code infilling, fine-tuned on up to 16K tokens and handles up to 100K tokens at inference.
• ConvNeXt Meta AI · Vision/CV · 2022 — Meta AI's ConvNeXt, a family of pure convolutional networks built by gradually modernizing a standard ResNet toward Vision Transformer design choices while using only standard ConvNet modules; ConvNeXt reached 87.8% ImageNet top-1 accuracy and beat Swin Transformers on COCO detection and ADE20K segmentation, keeping the simplicity and efficiency of standard ConvNets.
• DeiT Meta AI · Vision/CV · 2020 — Meta AI's DeiT (Data-efficient image Transformers), convolution-free vision transformers trained on ImageNet alone on a single computer in under 3 days; the reference 86M-parameter model reached 83.1% top-1 with no external data, and a transformer-specific distillation token, especially with a convnet teacher, pushed accuracy as high as 85.2%.
• Detectron2 Meta AI · Vision/CV · 2019 — Facebook AI Research's Detectron2, its next-generation library of state-of-the-art detection and segmentation algorithms, succeeding Detectron and maskrcnn-benchmark; it adds panoptic segmentation, DensePose, Cascade R-CNN, rotated boxes, PointRend, DeepLab, ViTDet and MViTv2, trains faster, exports models to TorchScript or Caffe2, and comes with a model zoo of trained baselines.
• DINO Meta AI · Vision/CV · 2021 — Meta AI's DINO, a self-supervised training method for Vision Transformers that works as self-distillation with no labels; its ViT features contain explicit semantic-segmentation information and make strong k-NN classifiers (78.3% ImageNet top-1 with a small ViT), and a ViT-Base reached 80.1% top-1 in linear evaluation.
• Emu Meta AI · Image Generation · 2023 — Meta's Emu text-to-image model: a latent diffusion model pretrained on 1.1 billion image-text pairs, then quality-tuned on a few thousand carefully selected, highly aesthetic images; Meta reports it won 82.9% of comparisons against its pretrained-only version and was preferred over SDXL 1.0 for visual appeal 68.4% and 71.3% of the time on two benchmarks.
• Emu Edit Meta AI · Image Editing · 2023 — Meta's Emu Edit (November 2023), a research model that edits images from plain-language instructions, trained to multitask across region-based and free-form edits and computer vision tasks such as detection and segmentation, all framed as generation, with learned task embeddings that steer it to the right kind of edit; it can pick up new tasks such as super-resolution from a few examples, and Meta released a seven-task editing benchmark along with its outputs.
• EnCodec Meta AI · Audio Codec · 2022 — Meta's EnCodec, a high-fidelity neural audio codec released as two multi-bandwidth models: a causal 24 kHz mono model trained on varied audio (1.5 to 24 kbps) and a non-causal 48 kHz stereo model trained on music (3 to 24 kbps), each paired with a language model that can compress the representation by up to 40% more without further quality loss.
• ESM-2 Meta AI · Science/Biology · 2022 — Meta FAIR's ESM-2 transformer protein language models, in six sizes from 8M to 15B parameters, which predict protein structure, function and other properties directly from individual sequences; Meta reported they outperformed all tested single-sequence protein language models on structure prediction tasks.
• ESMFold Meta AI · Science/Biology · 2022 — Meta FAIR's ESMFold, an end-to-end protein structure predictor that runs on the ESM-2 language model (about 690M parameters on top of a 3B-parameter ESM-2) and predicts 3D structure directly from a single protein sequence; released November 2022.
• ImageBind Meta AI · Multimodal · 2023 — Meta's ImageBind (May 9, 2023), an open-source model that learns one embedding space for six modalities, images and video, text, audio, depth, thermal and inertial (IMU) data, using only image-paired data rather than every combination of modalities; Meta reports new state-of-the-art emergent zero-shot recognition across modalities, beating specialist supervised models, and it enables cross-modal retrieval, embedding arithmetic and audio-to-image generation.
• Large Concept Model Meta AI · LLM · 2024 — Research architecture that models language over sentence-level 'concepts' rather than tokens, with concepts meant to be language- and modality-agnostic.
• Llama Meta AI · LLM · 2023 — Meta's LLaMA (Large Language Model Meta AI, February 24, 2023), foundation language models at 7B, 13B, 33B and 65B parameters trained only on publicly available data, 1.4 trillion tokens for the 33B and 65B models and 1 trillion for 7B, drawn from the 20 most-spoken languages; LLaMA-13B beat GPT-3 (175B) on most benchmarks and LLaMA-65B was competitive with Chinchilla-70B and PaLM-540B. Meta released it to researchers case by case under a noncommercial license.
• Llama 2 Meta AI · LLM · 2023 — Meta's Llama 2 family: pretrained and chat-tuned (Llama-2-Chat) text models at 7B, 13B and 70B parameters, with a 4K context and 2T pretraining tokens; the 70B model uses grouped-query attention. Meta reported Llama-2-Chat beat open-source chat models on most benchmarks it tested and was on par with some closed models, such as ChatGPT and PaLM, in its human evaluations.
• Llama 2 13B Meta AI · LLM · 2023 — 13B-parameter model in Meta's Llama 2 family, in pretrained and chat-tuned (Llama-2-Chat) versions: text only, with a 4K context and 2T pretraining tokens.
• Llama 2 70B Meta AI · LLM · 2023 — 70B-parameter model in Meta's Llama 2 family, in pretrained and chat-tuned (Llama-2-Chat) versions: text only, with a 4K context, 2T pretraining tokens and grouped-query attention. Meta reported the chat version was on par with some closed models, such as ChatGPT and PaLM, in its human evaluations of helpfulness and safety.
• Llama 2 7B Meta AI · LLM · 2023 — 7B-parameter model in Meta's Llama 2 family, in pretrained and chat-tuned (Llama-2-Chat) versions: text only, with a 4K context and 2T pretraining tokens.
• Llama 2 Chat 70B Meta AI · LLM · 2023 — Chat-tuned 70B Llama 2 model (Llama-2-Chat), optimized for dialogue with SFT and RLHF: text only, with a 4K context, 2T pretraining tokens and grouped-query attention. Meta reported Llama-2-Chat beat open-source chat models on most benchmarks it tested and was on par with some closed models, such as ChatGPT and PaLM, in its human evaluations.
• Llama 3 Meta AI · LLM · 2024 — Meta's Llama 3 (April 18, 2024), pretrained and instruction-tuned 8B and 70B open-weight language models: decoder-only transformers with a 128K-token vocabulary and grouped query attention, pretrained on over 15T tokens of public data, seven times Llama 2's dataset with four times more code and over 5% non-English text in 30+ languages. Meta called them the best open source models of their class, said 400B+ versions were still training, and released them under the Meta Llama 3 Community License.
• Llama 3 70B Meta AI · LLM · 2024 — 70B-parameter model in Meta's Llama 3 family, in pretrained and instruction-tuned versions: text only, with an 8K context, a 128K-token vocabulary and grouped-query attention, pretrained on 15T+ tokens of public online data (knowledge cutoff December 2023); the instruction-tuned version uses SFT and RLHF.
• Llama 3 8B Meta AI · LLM · 2024 — 8B-parameter model in Meta's Llama 3 family, in pretrained and instruction-tuned versions: text only, with an 8K context, a 128K-token vocabulary and grouped-query attention, pretrained on 15T+ tokens of public online data (knowledge cutoff March 2023); the instruction-tuned version uses SFT and RLHF.
• Llama 3.1 Meta AI · LLM · 2024 — Meta's Llama 3.1 family: 8B, 70B and 405B text models with a 128K context, pretrained on 15T+ tokens (knowledge cutoff December 2023) and officially supporting eight languages (English, German, French, Italian, Portuguese, Hindi, Spanish and Thai); the instruction-tuned versions are optimized for multilingual dialogue.
• Llama 3.1 70B Meta AI · LLM · 2024 — 70B-parameter model in Meta's Llama 3.1 family, in pretrained and instruction-tuned versions: multilingual text in, text and code out, with a 128K context and grouped-query attention, pretrained on 15T+ tokens (knowledge cutoff December 2023); officially supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• Llama 3.1 8B Meta AI · LLM · 2024 — 8B-parameter model in Meta's Llama 3.1 family, in pretrained and instruction-tuned versions: multilingual text in, text and code out, with a 128K context and grouped-query attention, pretrained on 15T+ tokens (knowledge cutoff December 2023); officially supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• Llama 3.1-405B Meta AI · LLM · 2024 — Largest model in Meta's Llama 3.1 family (405B parameters), in pretrained and instruction-tuned versions: multilingual text in, text and code out, with a 128K context and grouped-query attention, pretrained on 15T+ tokens (knowledge cutoff December 2023); officially supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• Llama 3.2 11B Vision Meta AI · Multimodal LLM · 2024 — 11B (10.6B-parameter) image-reasoning model in Meta's Llama 3.2-Vision family: takes text and images and returns text, with a 128K context. It adds a separately trained vision adapter (cross-attention layers) to the Llama 3.1 text model and was trained on 6B image-text pairs; for image-plus-text tasks only English is officially supported.
• Llama 3.2 1B Meta AI · LLM · 2024 — 1B (1.23B-parameter) text model in Meta's Llama 3.2 family, instruction-tuned for multilingual dialogue including agentic retrieval and summarization: 128K context (8K in the quantized versions), pretrained on up to 9T tokens (knowledge cutoff December 2023); officially supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• Llama 3.2 3B Meta AI · LLM · 2024 — 3B (3.21B-parameter) text model in Meta's Llama 3.2 family, instruction-tuned for multilingual dialogue including agentic retrieval and summarization: 128K context (8K in the quantized versions), pretrained on up to 9T tokens (knowledge cutoff December 2023); officially supports English, German, French, Italian, Portuguese, Hindi, Spanish and Thai.
• Llama 3.2 90B Vision Meta AI · Multimodal LLM · 2024 — The largest Llama 3.2 model (90B), adding image understanding to Llama's text model (September 2024); Meta's Llama 4 models (April 2025) have since succeeded it.
• Llama 3.3 Meta AI · LLM · 2024 — 70B text-only, instruction-tuned multilingual model with a 128K context window, trained on 15T+ tokens. Open weights (Llama 3.3 Community License).
• Llama 4 Meta AI · LLM · 2025 — Meta's Llama 4 herd (April 5, 2025): natively multimodal models and Meta's first built with mixture-of-experts, using early fusion of text and vision tokens and pre-trained on over 30 trillion tokens in 200 languages. Open weights came for Llama 4 Scout (17B active parameters, 16 experts, 109B total, 10M-token context) and Llama 4 Maverick (17B active, 128 experts, 400B total); Meta credits their quality to distilling from Llama 4 Behemoth, a 288B-active-parameter teacher still in training.
• Llama 4 Maverick Meta AI · LLM · 2025 — Natively multimodal mixture-of-experts model: 17B active / 400B total parameters, with 128 routed experts plus a shared expert. Meta says it beats GPT-4o and Gemini 2.0 Flash on widely reported benchmarks. Open weights.
• Llama 4 Scout Meta AI · LLM (MoE) · 2025 — Meta's Llama 4 Scout (April 5, 2025), an open-weight, natively multimodal mixture-of-experts model with 17B active parameters, 16 experts and 109B total that fits on a single NVIDIA H100 GPU with Int4 quantization. Pre-trained and post-trained at 256K context, it supports a context window of 10 million tokens, up from Llama 3's 128K, and Meta reports better results than Gemma 3, Gemini 2.0 Flash-Lite and Mistral 3.1 across widely reported benchmarks.
• Llama Guard Meta AI · Safety Model · 2023 — Meta's Llama Guard (December 2023), a 7B Llama 2-based safeguard model that labels both user prompts and model responses safe or unsafe and names the violated categories of its six-part taxonomy (violence and hate, sexual content, guns and illegal weapons, controlled substances, suicide and self-harm, criminal planning); tuned on about 13K examples, it matched or beat existing moderation tools on the OpenAI Moderation and ToxicChat benchmarks, and its categories can be changed by prompting.
• Llama Guard 2 Meta AI · Safety Model · 2024 — Meta's Llama Guard 2, an 8B Llama 3-based safeguard model that labels user prompts and model responses safe or unsafe across 11 hazard categories drawn from the MLCommons taxonomy and names the categories violated; on Meta's internal test set it scored an F1 of 0.915, against 0.665 for the original Llama Guard, 0.796 for GPT-4 and 0.347 for the OpenAI Moderation API, while keeping a low false-positive rate.
• Llama Guard 3 Meta AI · Safety Model · 2024 — Meta's Llama 3-based safety classifier for screening prompts and responses, released in 8B text, 1B and 11B vision versions; since succeeded by Llama Guard 4 (April 2025).
• Llama-3-Swallow-70B Meta AI · LLM · 2024 — Llama-3-Swallow-70B (July 1, 2024), a Japanese-enhanced base model from the Swallow project of Tokyo Institute of Technology's Okazaki and Yokota labs and AIST, continually pre-trained from Meta's Llama 3 mainly on Japanese data (the Swallow Corpus, Japanese Wikipedia and Laboro ParaCorpus) plus English and math text; it lifts the Japanese task average to 0.593 from Llama 3 70B's 0.568, at a slightly lower English average (0.675 vs 0.691). Released under the Meta Llama 3 Community License.
• MAE (Masked Autoencoder) Meta AI · Vision/CV · 2021 — Meta AI's MAE (Masked Autoencoders), a self-supervised vision method that masks a high share of image patches (such as 75%) and reconstructs the missing pixels, using an asymmetric encoder that sees only the visible patches and a lightweight decoder; it trains large models 3x or more faster, and a vanilla ViT-Huge reached 87.8% ImageNet accuracy, the best among methods using only ImageNet-1K data.
• Make-A-Video Meta AI · Video Generation · 2022 — Meta's Make-A-Video (September 2022), a research text-to-video system built on text-to-image models that learns appearance from captioned images and motion from unlabeled video, needing no paired text-video data; spatiotemporal layers, frame interpolation and two super-resolution models make the clips, it can also animate a still image or fill in motion between two images, and Meta reported 3x better text faithfulness and 3x higher quality than the prior state of the art in user studies.
• Meta AI Assistant Meta AI · Chat Interface · 2024 — Meta AI, Meta's free AI assistant, first announced at Connect 2023; built with Llama 3 from April 2024, it works in feed, chats and search across Facebook, Instagram, WhatsApp and Messenger and on the meta.ai website, answers with real-time web information, and creates images as you type with its Imagine feature. In April 2026 Meta made Muse Spark, Meta Superintelligence Labs' natively multimodal reasoning model, available at meta.ai and in the Meta AI app.
• Meta Muse Spark Meta AI · Reasoning LLM · 2026 — Meta's first closed-weight frontier model (Meta Superintelligence Labs, A. Wang). Artificial Analysis Index 52 at launch (4th, behind GPT-5.4, Gemini 3.1 Pro, Opus 4.6); led HealthBench Hard 42.8.
• MMS (Massively Multilingual Speech) Meta AI · Speech Recognition · 2023 — Meta AI's Massively Multilingual Speech (MMS) project, which extends speech technology from about 100 languages to over 1,000: a single speech recognition model for over 1,100 languages, language identification for over 4,000, pretrained models covering over 1,400, and text-to-speech for over 1,100; checkpoints are available in Hugging Face Transformers.
• Movie Gen Meta AI · Video Generation · 2024 — Meta's Movie Gen (October 4, 2024), research media foundation models: a 30B-parameter transformer that generates 1080p HD video of up to 16 seconds at 16 fps from text, variants for personalized video from a person's photo and for precise text-driven video editing, and a 13B audio model that makes synced sound effects and music of up to 45 seconds. Trained on licensed and public data, it beat rival models in Meta's human evaluations; Meta shared it as research toward a possible future release.
• Muse Glimmer Meta AI · Multimodal LLM · 2026 — Meta's compact open multimodal model: 29.6B dense parameters, Apache 2.0, 131K context - sized to run on a single consumer GPU.
• Muse Image Meta AI · Image Generation · 2026 — Meta Superintelligence Labs' first image model: agentic generation that calls web search and code execution mid-generation, multi-reference composition, Instagram social-context conditioning. Debuted #2 on Arena text-to-image (~1280).
• Muse Spark 1.1 Meta AI · LLM · 2026 — 1M-token-context agentic model rivaling GPT-5.5 and Opus 4.8 on agentic evals; claims #1 on MCP Atlas, JobBench, Humanity's Last Exam and Finance Agent V2; 20% vs Fable 5's 11% on the held-back Vals AI Harvey legal-agent benchmark. Computer use across desktop/browser/mobile with parallel subagent delegation. Replit, Cline and Box early partners. No open weights — a notable shift for Meta. Marks Meta's return to the frontier ('from three labs to five').
• Muse Spark 1.2 Meta AI · LLM · 2026 — Powers Muse Code, MSL's terminal coding agent (beta, global): plans, writes and validates changes across large repos; early testing puts it around Grok 4.5 level using ~50% more tokens; 1M context. The 12-21x-cheaper contributor tier trades data for price.
• Muse Spark 1.3 Meta AI · LLM · 2026 — Agentic update to Meta's Muse Spark: ~20% fewer tool calls than 1.2 for equivalent tasks; blended cost around $0.10/M on the contributor tier. Powers the Muse Code terminal coding agent.
• Muse Video Meta AI · Video Generation · 2026 — Meta Superintelligence Labs' Muse Video, previewed July 7, 2026 alongside Muse Image as the lab's first media generation models; built on the same pretraining base as Muse Image, it generates video with native audio, and Meta reports competitive prompt adherence, visual fidelity and temporal consistency, ranking No. 3 for text-to-video on Arena (July 5, 2026). Meta named audio-video sync and fast, physically accurate motion as gaps and said it was coming soon to creators and Meta AI.
• MusicGen Meta AI · Music Generation · 2023 — Meta's MusicGen, a single-stage autoregressive transformer that generates music from text (and, in melody versions, from text plus a melody) over a 32 kHz EnCodec tokenizer; trained on 20K hours of licensed music and released in 300M, 1.5B and 3.3B sizes, plus stereo versions.
• NLLB-200 Meta AI · Translation · 2022 — Meta AI's NLLB-200 (No Language Left Behind), a machine-translation model for single-sentence translation among 200 languages, aimed mainly at research, especially on low-resource languages; Meta says it is a research model not released for production, not meant for domain-specific (such as medical or legal) or document translation, and trained on inputs of up to 512 tokens.
• Nougat Meta AI · Document AI · 2023 — Meta AI's Nougat (Neural Optical Understanding for Academic Documents), a model that parses academic PDFs into Mathpix-compatible markdown, including LaTeX math and tables; available in small and base versions with a command-line tool and an API.
• OPT-175B Meta AI · LLM · 2022 — Meta AI's OPT-175B, the largest of the Open Pre-trained Transformers (OPT), a suite of decoder-only language models from 125M to 175B parameters trained to roughly match the GPT-3 class in size and performance; the smaller models are openly downloadable, while OPT-175B's weights are available to researchers on request.
• RoBERTa Meta AI · LLM · 2019 — Meta AI's RoBERTa (Robustly Optimized BERT Pretraining Approach), from a replication study of BERT pretraining that measured the impact of key hyperparameters and training-data size and found BERT was significantly undertrained; the resulting model set state-of-the-art results on GLUE, RACE and SQuAD, and the models and code were released.
• SAM (Segment Anything) Meta AI · Vision/CV · 2023 — Meta AI's Segment Anything Model (SAM), which produces high-quality object masks from prompts such as points or boxes, or masks for every object in an image; trained on 11 million images and 1.1 billion masks, with strong zero-shot performance across segmentation tasks. Meta has since released SAM 2, which extends it to video.
• SAM 2 Meta AI · Vision (Segmentation) · 2024 — Meta's Segment Anything Model 2 (SAM 2, July 29, 2024), the first unified model for real-time, promptable object segmentation in images and videos, using memory to follow objects across frames at about 44 frames per second; it needs about 3x fewer interactions than earlier video methods, beats SAM on images while being 6x faster, and segments unseen objects zero-shot. Code and weights are Apache 2.0, with the SA-V dataset (about 51,000 videos); SAM 2.1 followed in fall 2024.
• SAM 3 Meta AI · Vision/CV · 2025 — Meta's unified foundation model for promptable segmentation in images and video: it detects, segments and tracks objects from text prompts or visual prompts such as points, boxes and masks, and can segment every instance of a concept named in a short phrase. SAM 3.1 checkpoints followed in March 2026.
• SAM 3D Meta AI · Vision (3D Segmentation) · 2025 — Meta's SAM 3D (November 2025), two models that reconstruct 3D from a single natural image: SAM 3D Objects recovers the shape, texture and layout of selected objects and scenes, and SAM 3D Body estimates human body pose and shape, promptable with masks and 2D keypoints. Trained with a data engine that annotated almost 1 million images, SAM 3D Objects won at least 5:1 in human preference tests; Meta released checkpoints and inference code, and the models power Facebook Marketplace's View in Room.
• Seamless M4T Meta AI · Translation · 2023 — Meta's SeamlessM4T, a massively multilingual and multimodal translation model for about 100 languages, handling speech-to-speech, speech-to-text, text-to-speech and text-to-text translation plus speech recognition; v2, on the UnitY2 architecture, improved quality and speech-generation latency and underpins SeamlessExpressive and SeamlessStreaming.
• Segment Anything Model (SAM) Meta AI · Vision/CV · 2023 — Universal image segmentation. Click any point in any image and it segments the object. Promptable computer vision. 2023.
• V-JEPA 2 Meta AI · Vision/CV · 2025 — Meta's 1.2B-parameter self-supervised video world model (Joint Embedding Predictive Architecture) for understanding, predicting and planning in the physical world; supports robot planning with limited domain-specific training and shipped with new physical-reasoning benchmarks.
• Voicebox Meta AI · Speech Synthesis · 2023 — Meta's Voicebox (June 16, 2023), a non-autoregressive speech generation model built on Flow Matching that learns to infill speech from audio context and text, trained on over 50,000 hours of public domain audiobooks in six languages; it handles zero-shot and cross-lingual text-to-speech, noise removal, content editing and style conversion, and beat VALL-E on zero-shot TTS (1.9% vs 5.9% word error rate) while running up to 20x faster. Meta did not release the model or code, citing misuse risks.
• MetaVoice-1B MetaVoice · Speech Synthesis · 2024 — MetaVoice-1B, a 1.2B-parameter text-to-speech base model trained on 100K hours of speech, built for emotional speech rhythm and tone in English, zero-shot cloning of American and British voices from 30 seconds of reference audio, cross-lingual voice cloning with fine-tuning, and text of any length; released under Apache 2.0.
• BEiT Microsoft · Vision/CV · 2021 — Microsoft Research's BEiT (Bidirectional Encoder representation from Image Transformers, June 2021), which brought BERT-style masked pre-training to vision: images are split into patches and discrete visual tokens, some patches are masked, and the Vision Transformer learns to predict the missing tokens; the 86M-parameter base model reached 83.2% ImageNet-1K top-1 versus 81.8% for DeiT trained from scratch, and the 304M large model 86.3% using ImageNet-1K alone. Followed by BEiT v2 and BEiT-3.
• Bing Chat / Copilot Microsoft · Search AI · 2023 — Bing Chat, the AI chat Microsoft built into its Bing search engine, which by late 2023 ran on OpenAI models including GPT-4 and DALL-E 3 for text and images; on November 15, 2023, after more than 1 billion prompts and queries, Microsoft renamed Bing Chat and Bing Chat Enterprise Copilot, opened copilot.microsoft.com as its dedicated home and kept a combined search-and-chat experience on Bing.com. It continues as the free Microsoft Copilot.
• Bing Copilot Microsoft · Search AI · 2023 — Copilot in Bing, the AI chat built into Microsoft's Bing search engine: it launched as Bing Chat and was renamed Copilot on November 15, 2023, when Microsoft opened copilot.microsoft.com as a dedicated chat home while Bing.com kept a combined search-and-chat experience; by then it had handled more than 1 billion prompts and queries on OpenAI models including GPT-4 and DALL-E 3. The same assistant continues as the free Microsoft Copilot.
• BioGPT Microsoft · Science/Biology · 2022 — Microsoft Research's BioGPT (October 2022), a domain-specific generative Transformer language model pre-trained on large-scale biomedical literature, built because BERT-style biomedical models such as BioBERT and PubMedBERT cannot generate text; it outperformed previous models on most of six biomedical NLP tasks, setting a record 78.2% accuracy on PubMedQA and F1 scores of 44.98% on BC5CDR, 38.42% on KD-DTI and 40.76% on DDI relation extraction. MIT license.
• BitNet b1.58 2B4T Microsoft · LLM · 2025 — First open-source native 1-bit LLM at the 2B scale, trained on 4T tokens; Microsoft Research reports performance comparable to similar-size full-precision models with large memory, energy and latency savings. MIT license.
• Copilot (consumer) Microsoft · Chat Interface · 2024 — Microsoft Copilot, Microsoft's free AI assistant on the web, Windows, macOS, iOS, Android and in Edge: it chats, gives real-time answers from web search, helps with writing, brainstorming and summaries, creates images and summarizes web pages in Edge, and signing in adds chat history, longer conversations and voice. Microsoft 365 Personal, Family and Premium add Copilot in Word, Excel, PowerPoint, Outlook and OneNote, and Premium adds the Researcher, Analyst and Tasks agents.
• Copilot App Microsoft · Chat Interface · 2023 — The Microsoft Copilot app, Microsoft's standalone AI assistant app for Windows, macOS, iOS and Android, free to use: it chats, answers from real-time web search, helps write and summarize, creates images and supports voice conversations, with chat history and longer conversations for signed-in users. Work accounts get Copilot Chat in the same app with enterprise data protection, and the Microsoft 365 Copilot add-on grounds answers in company files, email and meetings.
• DeBERTa v3 Microsoft · LLM · 2021 — Microsoft's DeBERTaV3 (November 2021), which improves DeBERTa by replacing masked language modeling with ELECTRA-style replaced token detection and adding gradient-disentangled embedding sharing to avoid the 'tug-of-war' between generator and discriminator; DeBERTaV3 Large averaged 91.37% on GLUE, 1.37 points above DeBERTa and 1.91 above ELECTRA, and the multilingual mDeBERTa Base reached 79.8% zero-shot XNLI accuracy. Weights and code are released under the MIT license.
• E5 Microsoft · Embeddings · 2022 — Microsoft's E5 (December 2022), a family of general-purpose text embedding models trained contrastively on CCPairs, a curated large-scale dataset of weakly supervised text pairs, for retrieval, clustering and classification; zero-shot, it was the first model to beat the BM25 baseline on the BEIR retrieval benchmark without labeled data, and fine-tuned it topped MTEB, beating embedding models with 40x more parameters. Its weights are on Hugging Face under the MIT license.
• E5-Mistral-7B Microsoft · Embeddings · 2023 — Microsoft's E5-Mistral-7B (December 2023), a text embedding model initialized from Mistral-7B-v0.1 and fine-tuned with a standard contrastive loss, in under 1,000 training steps, on synthetic data that proprietary LLMs generated for hundreds of thousands of embedding tasks in 93 languages; trained on a mix of synthetic and labeled data, it set new state-of-the-art results on BEIR and MTEB. It produces 4,096-dimension embeddings, is recommended for English, and is MIT-licensed.
• Florence-2 Microsoft · Vision/CV · 2023 — Microsoft's Florence-2 (November 2023), a vision foundation model with a unified, prompt-based sequence-to-sequence design that takes text prompts as task instructions and returns text for captioning, object detection, grounding, segmentation and OCR; it was trained on FLD-5B, a co-developed dataset of 5.4 billion visual annotations on 126 million images, with strong zero-shot and fine-tuning results. Base (0.23B) and large (0.77B) versions are on Hugging Face under MIT.
• Kosmos-2 Microsoft · Multimodal LLM · 2023 — Microsoft's Kosmos-2, a multimodal large language model that can perceive object descriptions such as bounding boxes and ground text in images, writing referring expressions as Markdown-style links to location tokens; trained on GrIT, a large set of grounded image-text pairs, it handles phrase grounding, referring-expression comprehension and generation, perception-language tasks and language tasks, with code and weights released.
• LayoutLM v3 Microsoft · Document AI · 2022 — Microsoft's LayoutLMv3 (April 2022), a multimodal Transformer pre-trained for Document AI with unified text and image masking plus a word-patch alignment objective, giving one general-purpose model for text-centric tasks such as form and receipt understanding and document visual question answering and image-centric tasks such as document image classification and layout analysis, with state-of-the-art results on both at release. Weights are under CC BY-NC-SA 4.0.
• LinkedIn AI Writing Microsoft · AI Writing · 2023 — LinkedIn's AI-powered writing tool for drafting posts. LinkedIn's help page for the tool now says it is no longer available, and that LinkedIn will introduce new tooling to help improve post writing in the near future.
• MAI-1-preview Microsoft · LLM · 2025 — Microsoft AI's first end-to-end in-house foundation model: a mixture-of-experts LLM pre- and post-trained on ~15,000 NVIDIA H100 GPUs, opened for public testing on LMArena on 2025-08-28 with API access for trusted testers and a planned rollout to select Copilot text use cases.
• MAI-Code-1-Flash Microsoft · Code LLM · 2026 — MAI-Code-1-Flash (June 2, 2026), Microsoft AI's inference-efficient agentic coding model with 5B active parameters, trained without distillation inside the GitHub Copilot harness for everyday coding with adaptive thinking and strong instruction following: Microsoft reports 51.2% on SWE-Bench Pro against 35.2% for Claude Haiku 4.5, using up to 60% fewer tokens on harder problems. It rolled out to GitHub Copilot users in VS Code, and Microsoft AI now lists MAI-Code-1.1-Flash as a newer model.
• MAI-Cyber-1-Flash Microsoft · LLM · 2026 — Microsoft's first in-house cyber model: compact ~5B-active specialist paired with MDASH (orchestrating 100+ specialized agents); 95.95% CyberGym — about 12 points above Anthropic's Mythos — at roughly half prior cost, routing ~90% of detection work to the cheap specialist and escalating hard cases. Found 16 real Windows CVEs (4 critical RCEs); 21/21 on planted bugs with zero false positives.
• MAI-DS-R1 Microsoft · LLM · 2025 — DeepSeek-R1 post-trained by Microsoft AI to respond on previously blocked topics and improve its risk profile while keeping R1's reasoning performance. Open weights (MIT).
• MAI-Image-2.5 Microsoft · Image Generation · 2026 — MAI-Image-2.5 (June 2, 2026), Microsoft AI's image generation and editing model, with a faster, cheaper MAI-Image-2.5-Flash: it improves text rendering, product imagery and prompt adherence, makes localized edits that keep perspective, shadows and faces, and ranked No. 2 on Arena's image-edit leaderboard and No. 3 for text-to-image, ahead of GPT-Image-1.5 and Nano Banana Pro 2K. On Foundry it costs $5/$47 per million text-input/image-output tokens; it also powers PowerPoint and OneDrive.
• MAI-Image-2.5-Pro Microsoft · Image Generation · 2026 — Microsoft AI's MAI-Image-2.5-Pro (public preview July 23, 2026), which Microsoft calls its highest-fidelity image model to date, built for work where quality comes first, such as hero imagery, detailed editing and precise in-image text rendering; it sits alongside Microsoft's other MAI production models in Microsoft Foundry and the MAI Playground, priced at $5 per million text input tokens, $8 per million image input tokens and $106 per million image output tokens.
• MAI-Thinking-1 Microsoft · Reasoning LLM · 2026 — Microsoft AI's flagship reasoning model from the seven-model Build 2026 launch: 1T total / 35B active MoE trained from scratch on 33T tokens without distillation; matches leading models on key software-engineering benchmarks; preferred to Sonnet 4.6 in blind human side-by-sides. Marks Microsoft becoming a frontier lab in its own right rather than only an OpenAI channel.
• MAI-Transcribe-1.5 Microsoft · Speech Recognition · 2026 — MAI-Transcribe-1.5 (June 2, 2026), Microsoft AI's speech-to-text model, expanded from 25 to 43 languages with best-in-class word error rates on FLEURS and a 2.4% WER that ranks third on Artificial Analysis's leaderboard; keyword biasing for domain terms cuts errors by 30%, it is tuned for noisy audio and runs up to five times faster than rival models on long audio. It powers Copilot, Teams, GitHub and Dynamics 365 Contact Center and is on Microsoft Foundry; diarization and streaming are planned.
• MAI-Voice-1 Microsoft · Voice / TTS · 2025 — Microsoft AI's first in-house expressive speech-generation model: generates a minute of audio in under a second on a single GPU; powers Copilot Daily and Podcasts and can be tried in Copilot Labs.
• MAI-Voice-2 Microsoft · Voice / TTS · 2026 — MAI-Voice-2 (June 2, 2026), Microsoft AI's most expressive text-to-speech model, expanding from English only to 15 languages: it creates custom voices from a few seconds of reference audio (5 to 60 seconds recommended) without retraining, with consent enforced so only authorized, licensed voices are used. Listeners preferred it to MAI-Voice-1 in 72.1% of 2,500 tests and to human recordings 45.5% to 44%; it is in Microsoft Foundry and coming to VS Code and Dynamics 365 Contact Center.
• MAI-Voice-2-Flash Microsoft · Voice / TTS · 2026 — Microsoft AI's MAI-Voice-2-Flash (public preview July 23, 2026, first shown at Build), the fast, efficient tier of its MAI voice models for high-volume voice experiences, keeping MAI-Voice-2's natural prosody and acoustic quality while running 2x faster at 32% lower cost, $15 per million characters; it powers Dynamics 365 Contact Center, where Microsoft says it cut GPU costs by up to 89%, and backs speech-to-speech voice agents built in Voice Live.
• Microsoft Copilot for M365 Microsoft · Productivity AI · 2023 — Microsoft 365 Copilot, Microsoft's paid business add-on that brings full, priority-access AI into Word, Excel, PowerPoint, Outlook, OneNote and Teams and grounds answers in an organization's emails, files, meetings and calendars through Microsoft Graph and Work IQ, within each user's permissions. Copilot Chat on web data is included with eligible Microsoft 365 plans, usage-based billing adds Copilot Cowork, Copilot Code and agents over company data, and some experiences use Anthropic models.
• Microsoft Mu Microsoft · LLM · 2025 — 330M-parameter on-device encoder-decoder language model built for the NPUs in Copilot+ PCs; it powers the agent in Windows Settings.
• Microsoft phi-4 Microsoft · LLM · 2024 — Microsoft's Phi-4 (December 12, 2024), a 14B-parameter dense decoder-only model built with a training recipe centered on data quality that weaves synthetic data throughout training; trained on 9.8T tokens over 21 days on 1,920 H100 GPUs, it surpasses its teacher model GPT-4 on STEM-focused question answering and performs strongly for its size on reasoning-focused benchmarks, with a 16K-token context and an MIT license.
• Microsoft Security Copilot Microsoft · Security AI · 2023 — Microsoft Security Copilot, Microsoft's generative AI assistant for security teams, helps in natural language with incident response, threat hunting, intelligence gathering and posture management, writes KQL queries and analyzes suspicious scripts, drawing on plugins and threat intelligence, standalone or inside Defender XDR, Sentinel, Intune, Entra and tools like ServiceNow and Jamf. Since November 2025 it comes with Microsoft 365 E5, along with 40 new Microsoft and partner agents.
• NaturalSpeech 3 Microsoft · Speech Synthesis · 2024 — Microsoft's NaturalSpeech 3, a zero-shot text-to-speech research system that uses a neural codec with factorized vector quantization to split speech into content, prosody, timbre and acoustic-detail subspaces, then generates each with a factorized diffusion model; Microsoft reports it beat state-of-the-art TTS systems on quality, similarity, prosody and intelligibility, matched human recordings in quality, and improved further at 1B parameters and 200K hours of training data.
• Orca Microsoft · LLM · 2023 — Microsoft Research's Orca (June 2023), a 13B-parameter model that learns to imitate the reasoning process of large foundation models from rich GPT-4 signals, including explanation traces and step-by-step thought processes, with ChatGPT as a teacher assistant; it beat Vicuna-13B by more than 100% on Big-Bench Hard and 42% on AGIEval, reached parity with ChatGPT on BBH, and was competitive on professional and academic exams such as the SAT, LSAT, GRE and GMAT.
• Orca 2 Microsoft · LLM · 2023 — Microsoft Research's Orca 2 (November 2023), 7B and 13B research models built on Llama 2 and trained on synthetic data that teaches small models several reasoning strategies, such as step-by-step, recall-then-generate and direct answer, and when to use each; across 15 benchmarks it significantly surpassed models of similar size and matched or beat models 5-10x larger on complex zero-shot reasoning. It is not optimized for chat, and its weights are released for research.
• Phi-1 Microsoft · LLM · 2023 — Microsoft's Phi-1, a 1.3B-parameter transformer specialized for basic Python coding, trained on Python code from The Stack v1.2, StackOverflow Q&A, code_contests and synthetic Python textbooks and exercises generated by GPT-3.5; a research model not tested for production coding.
• Phi-1.5 Microsoft · LLM · 2023 — Microsoft's Phi-1.5, a 1.3B-parameter transformer trained on Phi-1's data sources plus new synthetic NLP texts, with generic web-crawl data left out for safety; Microsoft reports nearly state-of-the-art common-sense, language-understanding and logical-reasoning results among models under 10B parameters. A base model with no instruction tuning or RLHF, released for safety research.
• Phi-2 Microsoft · LLM · 2023 — Microsoft's Phi-2 (December 2023), a 2.7B-parameter Transformer trained on the same data sources as Phi-1.5 plus new synthetic NLP texts and filtered websites, which showed nearly state-of-the-art performance among models under 13B parameters on common-sense, language-understanding and logical-reasoning benchmarks; released without RLHF as an unrestricted small model for research on safety challenges, it is best suited to QA, chat and code prompts. MIT license.
• Phi-3 Microsoft · LLM · 2024 — Microsoft's Phi-3 family (April 2024), led by phi-3-mini, a 3.8B-parameter model trained on 3.3T tokens of heavily filtered web and synthetic data that rivals Mixtral 8x7B and GPT-3.5 (69% on MMLU, 8.38 on MT-Bench) while small enough to run on a phone; phi-3-small (7B) and phi-3-medium (14B) reached 75% and 78% on MMLU, and the later Phi-3.5 line added mini, MoE (6.6B active parameters) and Vision models. MIT license.
• Phi-3 Medium Microsoft · LLM · 2024 — Microsoft's 14B-parameter dense Phi-3 model, trained on 4.8T tokens and post-trained with SFT and DPO; released in 4K and 128K context versions.
• Phi-3 Mini Microsoft · LLM · 2024 — Microsoft's 3.8B-parameter dense Phi-3 model, trained on synthetic and filtered public web data chosen for reasoning-dense content and post-trained with SFT and DPO; released in 4K and 128K context versions.
• Phi-3 Small Microsoft · LLM · 2024 — Microsoft's 7B-parameter dense Phi-3 model, using alternating dense and block-sparse attention and trained on 4.8T tokens; released in 8K and 128K context versions.
• Phi-3 Vision Microsoft · Multimodal LLM · 2024 — Microsoft's 4.2B-parameter multimodal Phi-3 model (an image encoder plus the Phi-3 Mini language model) with a 128K context, trained on 500B vision and text tokens and best suited to single-image prompts.
• Phi-3.5 Microsoft · LLM · 2024 — Microsoft's Phi-3.5 series, which extends Phi-3 with stronger multilingual, multimodal and long-context abilities: Phi-3.5-mini, Phi-3.5-MoE (16 x 3.8B experts, 6.6B active parameters) and Phi-3.5-Vision (4.2B parameters, single- and multi-image input); Microsoft reports Phi-3.5-MoE beats similar-size open models such as Llama 3.1 and Mixtral on language reasoning, math and code and is on par with Gemini 1.5 Flash and GPT-4o-mini.
• Phi-3.5 Mini Microsoft · LLM · 2024 — Microsoft's 3.8B-parameter dense Phi-3.5 update to Phi-3 Mini, with a 128K context, 3.4T training tokens and post-training (SFT, PPO and DPO) aimed at multilingual, multi-turn and reasoning quality; supports 23 languages.
• Phi-3.5 MoE Microsoft · LLM · 2024 — Microsoft's mixture-of-experts Phi-3.5 model: 16 experts of 3.8B parameters with 6.6B active (2 experts per token), a 128K context and 4.9T training tokens; supports 23 languages.
• Phi-3.5 Vision Microsoft · Multimodal LLM · 2024 — Microsoft's 4.2B-parameter multimodal Phi-3.5 model with a 128K context that adds multi-frame input for detailed image comparison, multi-image summarization and video summarization; trained on 500B vision and text tokens.
• Phi-4 Microsoft · LLM · 2024 — Microsoft's Phi-4 (December 12, 2024), a 14B-parameter dense decoder-only model built with a training recipe centered on data quality that weaves synthetic data throughout training; trained on 9.8T tokens over 21 days on 1,920 H100 GPUs, it surpasses its teacher model GPT-4 on STEM-focused question answering and performs strongly for its size on reasoning-focused benchmarks, with a 16K-token context and an MIT license.
• Phi-4-mini Microsoft · LLM · 2025 — Microsoft's Phi-4-mini (February 2025), a 3.8B-parameter dense decoder-only open model trained on 5T tokens of high-quality web and synthetic data, which Microsoft says matches models twice its size on math and coding tasks requiring complex reasoning; compared with Phi-3.5-mini it adds a 200K-token vocabulary for multilingual use and grouped-query attention, with better instruction following and function calling, 23 supported languages and a 128K-token context. MIT license.
• Phi-4-mini-flash-reasoning Microsoft · Reasoning LLM · 2025 — Microsoft's Phi-4-mini-flash-reasoning, a 3.8B-parameter open model for multi-step math reasoning in memory-, compute- or latency-constrained settings, built on the SambaY decoder-hybrid-decoder architecture that mixes state-space layers, a gated memory unit and differential attention; with a 64K-token context it beat Phi-4-mini-reasoning on Math500, AIME 2024/25 and GPQA Diamond without reinforcement learning while decoding up to 10x faster on long generations. MIT license.
• Phi-4-multimodal Microsoft · Multimodal LLM · 2025 — Microsoft's Phi-4-multimodal (February 2025), a 5.6B-parameter open model that takes text, image and speech/audio input in a single network, built on Phi-4-mini with vision and speech encoders and modality-specific LoRA adapters and routers; trained on 5T text tokens, 2.3M hours of speech and 1.1T image-text tokens, it ranked first on the OpenASR leaderboard at release and beat larger vision-language and speech models on many tasks. 128K context, MIT license.
• Phi-4-reasoning-plus Microsoft · Reasoning LLM · 2025 — Microsoft's Phi-4-reasoning-plus (April 30, 2025), a 14B-parameter open reasoning model made by supervised fine-tuning of Phi-4 on curated 'teachable' prompts with reasoning traces from o3-mini, then a short phase of outcome-based reinforcement learning that produces longer reasoning chains; Microsoft says it outperforms much larger open models such as DeepSeek-R1-Distill-Llama-70B and approaches the full DeepSeek-R1 on math, science, coding and planning. 32K-token context; MIT license.
• PubMedBERT Microsoft · Science/Biology · 2020 — Microsoft Research's PubMedBERT, now named BiomedBERT (July 2020), a BERT model pretrained from scratch on PubMed abstracts (and, in one version, PubMed Central full-text articles), from research showing that in text-rich domains such as biomedicine, pretraining from scratch beats continual pretraining of general-domain models; it set state-of-the-art results across biomedical NLP and came with the BLURB benchmark and leaderboard. MIT license.
• ResNet-101 Microsoft · Vision/CV · 2015 — Microsoft Research's ResNet-101, a 101-layer residual network from the paper that introduced residual learning, in which layers learn residual functions relative to their inputs so much deeper networks train well; residual nets won first place in ILSVRC 2015 classification (3.57% ImageNet test error for an ensemble) and in ImageNet detection and localization and COCO detection and segmentation.
• ResNet-152 Microsoft · Vision/CV · 2015 — Microsoft Research's ResNet-152, the deepest ImageNet model (152 layers, 8x deeper than VGG nets yet lower in complexity) in the paper that introduced residual learning, in which layers learn residual functions relative to their inputs; residual nets won first place in ILSVRC 2015 classification (3.57% ImageNet test error for an ensemble) and in ImageNet detection and localization and COCO detection and segmentation.
• ResNet-50 Microsoft · Vision/CV · 2015 — ResNet-50, the 50-layer version of the residual networks Microsoft Research introduced in December 2015, which learn residual functions with reference to layer inputs so that far deeper networks become easier to optimize; an ensemble of residual nets up to 152 layers deep reached 3.57% error on ImageNet and won ILSVRC 2015 classification, and the networks also won ImageNet detection and localization and COCO detection and segmentation.
• Swin Transformer Microsoft · Vision/CV · 2021 — Microsoft's Swin Transformer, a hierarchical vision Transformer that computes self-attention within shifted, non-overlapping local windows, giving linear complexity in image size and making it a general-purpose vision backbone; it reached 87.3% ImageNet-1K top-1, 58.7 box AP and 51.1 mask AP on COCO, and 53.5 mIoU on ADE20K, well above the previous state of the art.
• VALL-E Microsoft · Speech Synthesis · 2023 — Microsoft's VALL-E, which treats text-to-speech as conditional language modeling over discrete codes from a neural audio codec; trained on 60K hours of English speech, it can synthesize personalized speech from a 3-second recording of an unseen speaker, preserving the speaker's emotion and acoustic environment, and Microsoft reports it beat the previous best zero-shot TTS system on naturalness and speaker similarity.
• VALL-E X Microsoft · Speech Synthesis · 2023 — Microsoft's VALL-E X, a cross-lingual neural codec language model extending VALL-E: prompted with one utterance in a source language plus target-language text, it generates target-language speech that keeps the unseen speaker's voice, emotion and acoustic environment, enabling zero-shot cross-lingual speech synthesis and speech-to-speech translation, with a language ID to control foreign accent.
• WizardLM Microsoft · LLM · 2023 — WizardLM, LLaMA fine-tuned on instruction data that an LLM generated with Evol-Instruct, which rewrites seed instructions step by step into more complex ones; its authors found Evol-Instruct data beat human-written instructions in human evaluation, and in GPT-4 evaluation WizardLM reached more than 90% of ChatGPT's capacity on 17 of 29 skills.
• WizardLM 2 Microsoft · LLM · 2024 — WizardLM-2, a family of open large language models (WizardLM-2 8x22B, 70B and 7B) that its team released on April 15, 2024, post-trained with a fully AI-powered synthetic-data system built on Evol-Instruct and RLEIF reinforcement learning for complex chat, multilingual, reasoning and agent tasks; the team called 8x22B the best open model in its internal evaluation on highly complex tasks and 7B comparable to open models 10 times its size. Its Hugging Face model page now returns not found.
• Megatron-Turing NLG Microsoft and Nvidia · LLM · 2022 — 530B-parameter transformer language model trained jointly by Microsoft and NVIDIA with DeepSpeed and Megatron 3D parallelism; the largest monolithic transformer language model at its release.
• WizardCoder Microsoft/Community · Code LLM · 2023 — WizardCoder, code LLMs fine-tuned with complex instructions by adapting the Evol-Instruct method to code; its authors reported it beat all other open-source code LLMs by a wide margin on HumanEval, HumanEval+, MBPP and DS-1000, and outperformed Anthropic's Claude and Google's Bard on HumanEval and HumanEval+ at release.
• Midjourney Alpha (Web) Midjourney · Image Generation · 2024 — Web-based Midjourney interface.
• Midjourney V1 Video Midjourney · Video Generation · 2025 — Midjourney's first video model: image-to-video that returns four 5-second clips per job, with automatic or manual motion prompts and low/high motion settings; clips can be extended about 4 seconds at a time to roughly 21 seconds.
• Midjourney v5 Midjourney · Image Generation · 2023 — Photorealistic image generation.
• Midjourney v5.2 Midjourney · Image Generation · 2023 — Improved aesthetics and coherence.
• Midjourney v6 Midjourney · Image Generation · 2023 — Midjourney's V6 image model, the default version from February 14 to July 30, 2024, with more accurate handling of longer prompts, improved coherence and knowledge, and better image prompting and remixing.
• Midjourney v6.1 Midjourney · Image Generation · 2024 — Midjourney's V6.1 image model, released July 30, 2024 and the default version until June 16, 2025: more coherent images with more precise details and textures, generated about 25% faster than V6.
• Midjourney v7 Midjourney · Image Generation · 2025 — Midjourney's V7 image model, released April 3, 2025 and the default version from June 17, 2025 to June 9, 2026: more precise handling of text and image prompts, richer textures and more coherent bodies, hands and objects; introduced Draft Mode and Omni Reference.
• Midjourney V8.1 Midjourney · Image Generation · 2026 — Released on midjourney.com Apr 30, 2026 as Midjourney's fastest model: standard jobs render about 4-5x faster than earlier versions, with better prompt reading and detail retention, Raw mode, and native 2K HD images without upscaling. Requires an unlocked V7/V8 personalization profile.
• Hailuo AI MiniMax · Video Generation · 2024 — Hailuo AI, MiniMax's AI video and image creation platform for web and mobile, which turns text or photos into videos with MiniMax's video models, including Hailuo 2.3, which improved body movement, stylization and facial micro-expressions over Hailuo 02 and added a cheaper 2.3 Fast version, and MiniMax H3, which adds native stereo sound; a Media Agent builds multi-modal videos in one click. Free users get watermarked downloads; paid plans add credits, watermark-free downloads and priority.
• MiniMax H3 (Hailuo 3.0) MiniMax · Video Generation · 2026 — MiniMax H3, MiniMax's open-weight omni-modal generation system: it understands combined text, image, video and audio context and generates video with native stereo sound at up to 2K and 15 seconds, using a 33B-parameter H3-Omni-Transformer, a Qwen3-VL-32B-based encoder, separate video and audio VAEs and a 2K regeneration step. MiniMax says its 2K price per second is under a third of mainstream models'. It runs in Hailuo AI and the MiniMax API; its license excludes the US, EU, UK and South Korea.
• MiniMax Talkie MiniMax · Chat Interface · 2024 — Talkie, MiniMax's AI character-chat app: users chat with thousands of pre-made AI characters across dozens of categories or create their own by setting a personality, scenario and greeting, and characters keep long-term memory and can exchange voice messages, with chats synced across devices and content moderated under community guidelines. Talkie is free to use, with optional Talkie+ perks.
• MiniMax-M1 MiniMax · LLM · 2025 — Open-weight hybrid-attention reasoning model: a mixture-of-experts with 456B total and 45.9B active parameters per token, lightning attention and a native 1M-token context, built on MiniMax-Text-01. Apache 2.0.
• MiniMax-M2 MiniMax · LLM · 2025 — MiniMax-M2, MiniMax's open-weight (modified MIT) mixture-of-experts model for coding and agents, with 230B total and 10B active parameters: it runs long tool chains across shell, browser and code runners, scores 69.4 on SWE-bench Verified and 46.3 on Terminal-Bench, and ranked first among open models on Artificial Analysis's composite at launch. MiniMax priced its API at $0.30/$1.20 per million tokens, which it called 8% of Claude Sonnet 4.5's price, and launched an M2-powered MiniMax Agent.
• MiniMax-M2.1 MiniMax · LLM · 2025 — Coding-focused open-weights update: stronger multi-language programming (Rust, Java, Go, C++, Kotlin, Objective-C, TS/JS), better web and native-mobile app development, interleaved thinking for composite office-style instructions, and more concise, lower-token outputs.
• MiniMax-M2.5 MiniMax · LLM · 2026 — MiniMax-M2.5, MiniMax's open-weight (modified MIT) coding and agent model, trained with RL in hundreds of thousands of real-world environments: MiniMax reports 80.2% on SWE-Bench Verified, 51.3% on Multi-SWE-Bench and 76.3% on BrowseComp, says it finishes SWE-Bench Verified 37% faster than M2.1 and improved its Word, PowerPoint and Excel work. A Lightning version runs at 100 tokens a second, which MiniMax says costs $1 an hour; M2.5 is now a legacy API model at $0.30/$1.20 per million tokens.
• MiniMax-M2.7 MiniMax · Multimodal LLM · 2026 — MiniMax-M2.7, which MiniMax calls its first model to take part in its own evolution: an internal version updated its own memory, built skills for reinforcement-learning experiments and ran 100+ rounds of autonomous scaffold optimization for a 30% gain. MiniMax reports 56.22% on SWE-Pro, 57.0% on Terminal Bench 2 and a 1495 Elo on GDPval-AA, the best among open models; it also supports native Agent Teams and Office-file editing. Its open weights bar commercial use without MiniMax's permission.
• MiniMax-M3 MiniMax · LLM · 2026 — MiniMax-M3, MiniMax's open-weight flagship (about 428B total and 23B active parameters), trained natively on mixed modalities from the first step: it takes image and video input and can operate a desktop computer. Its MiniMax Sparse Attention gives a 1M-token context with more than 9x faster prefill and 15x faster decoding than M2 at that length; MiniMax reports 59.0% on SWE-Bench Pro and 66.0% on Terminal-Bench 2.1, and it offers enabled, adaptive and disabled reasoning modes.
• MiniMax-Text-01 MiniMax · LLM · 2025 — MiniMax-Text-01 (January 2025), MiniMax's open-weight language model with 456B total and 45.9B active parameters: a 32-expert mixture-of-experts with one softmax attention layer after every seven Lightning Attention layers, trained on contexts up to 1 million tokens and handling up to 4 million at inference. MiniMax's paper says it matches GPT-4o and Claude 3.5 Sonnet with a 20 to 32 times longer context window; it supports function calling, and MiniMax-VL-01 is its vision-language sibling.
• Codestral Mistral AI · Code LLM · 2024 — Mistral AI's Codestral (May 29, 2024), its first code model: a 22B open-weight model trained on 80+ programming languages that completes functions, writes tests and fills in partial code with a fill-in-the-middle mechanism, and whose 32K-token context helped it top the RepoBench long-range code evaluation; released under the Mistral AI Non-Production License with commercial licenses on request, it had a free beta endpoint. This version left the API June 16, 2025.
• Codestral Mamba Mistral AI · Code LLM · 2024 — Mistral AI's Codestral Mamba (July 16, 2024), a 7.3B-parameter code model built on the Mamba2 architecture with help from Albert Gu and Tri Dao, offering linear-time inference and, in theory, unlimited sequence length; tested on in-context retrieval up to 256K tokens, it performs on par with state-of-the-art Transformer code models and was released under Apache 2.0 as a local code assistant. Mistral retired it from its API on June 6, 2025.
• Devstral 2 Mistral AI · Code LLM · 2025 — Mistral AI's Devstral 2 (December 9, 2025), a 123B-parameter dense coding model with a 256K-token context that scored 72.2% on SWE-bench Verified, which Mistral called the open state of the art for code agents and up to 7x more cost-efficient than Claude Sonnet on real-world tasks; released under a modified MIT license alongside the Apache 2.0 Devstral Small 2 (24B, 68.0%) and the Mistral Vibe CLI, it was retired from Mistral's API on July 31, 2026.
• Devstral Medium Mistral AI · Code LLM · 2025 — Mistral AI's Devstral Medium (July 10, 2025), an agentic coding model built with All Hands AI that scored 61.6% on SWE-bench Verified, which Mistral said surpassed Gemini 2.5 Pro and GPT-4.1 at a quarter of the price; offered through Mistral's API as devstral-medium-2507 at $0.40/$2 per million tokens, with private deployment and custom fine-tuning for enterprises, it launched with the open Devstral Small 1.1 and was retired from the API on May 31, 2026.
• Le Chat Mistral AI · Chat Interface · 2024 — Mistral AI's Le Chat (beta launched February 26, 2024), a multilingual conversational assistant that could run on Mistral Large, Mistral Small or the prototype Mistral Next, with tunable moderation and an Enterprise edition for self-deployment. On May 28, 2026 Le Chat became Vibe, one agent for work and coding, with a Work Mode for multi-step tasks across connected tools, a Code Mode for remote coding agents, and Free, Pro ($14.99 a month), Team and Enterprise plans.
• Magistral Medium Mistral AI · Reasoning LLM · 2025 — Mistral AI's Magistral Medium (June 10, 2025), the enterprise version of Magistral, Mistral's first reasoning model, released with the open 24B Magistral Small for domain-specific, transparent and multilingual reasoning whose chain of thought works across languages and alphabets; it scored 73.6% on AIME 2024 (90% with majority voting over 64) and launched in preview in Le Chat, the API and Amazon SageMaker. Its last version, 1.2, was retired from Mistral's API on July 31, 2026.
• Mathstral Mistral AI · LLM · 2024 — Mistral AI's Mathstral (July 16, 2024), a 7B model built on Mistral 7B that specializes in math and STEM reasoning, released for the science community as part of Mistral's collaboration with Project Numina; it scored 56.6% on MATH and 63.47% on MMLU, rising to 68.37% on MATH with majority voting and 74.59% with a reward model choosing among 64 candidates. Its instructed weights are on Hugging Face under Apache 2.0, and it has since been retired from Mistral's API.
• Ministral 3 Mistral AI · Multimodal LLM · 2025 — Small dense Mistral 3 models (14B, 8B, 3B) released Dec 2, 2025 alongside Mistral Large 3 under Apache 2.0; base, instruct and reasoning variants with native vision and multilingual support, aimed at edge and local deployment.
• Mistral 7B Mistral AI · LLM · 2023 — Mistral AI's Mistral 7B (September 27, 2023), which Mistral called the most powerful language model for its size to date: a 7.3B-parameter model that outperformed Llama 2 13B on all benchmarks and Llama 1 34B on many, approached CodeLlama 7B on code while staying good at English, and used grouped-query attention and sliding-window attention for faster, cheaper inference; released under Apache 2.0 with a chat fine-tune that beat Llama 2 13B chat. Retired from Mistral's API March 30, 2025.
• Mistral Devstral Medium Mistral AI · Code LLM · 2025 — Mistral AI's Devstral Medium (July 10, 2025), an agentic coding model built with All Hands AI that scored 61.6% on SWE-bench Verified, which Mistral said surpassed Gemini 2.5 Pro and GPT-4.1 at a quarter of the price; offered through Mistral's API as devstral-medium-2507 at $0.40/$2 per million tokens, with private deployment and custom fine-tuning for enterprises, it launched with the open Devstral Small 1.1 and was retired from the API on May 31, 2026.
• Mistral Embed Mistral AI · Embeddings · 2024 — Mistral AI's Mistral Embed, the embedding model behind the embedding endpoint Mistral opened in beta with La Plateforme on December 11, 2023; it produces 1,024-dimension text embeddings designed with retrieval in mind and scored 55.26 on MTEB retrieval at launch. Mistral still lists it, as version 23.12, among its current embedding models alongside Codestral Embed.
• Mistral Large Mistral AI · LLM · 2024 — Mistral's first flagship model (mistral-large-2402), natively fluent in English, French, Spanish, German and Italian, with a 32K-token context window; offered on la Plateforme and Azure from February 2024 until Mistral retired it from its API on June 16, 2025.
• Mistral Large 2 Mistral AI · LLM · 2024 — Mistral AI's Mistral Large 2 (July 24, 2024), a 123B-parameter model designed for single-node inference with a 128K-token context, dozens of languages and 80+ coding languages; trained on a large share of code, it performed on par with GPT-4o, Claude 3 Opus and Llama 3 405B on coding, scored 84.0% on MMLU as a pretrained model, and was trained to admit when it lacks a confident answer. Its weights came under the Mistral Research License; the API retired it on March 30, 2025.
• Mistral Large 3 Mistral AI · Multimodal LLM · 2025 — Mistral flagship, December 2025 generation (released Dec 2, 2025): sparse MoE with 41B active / 675B total parameters; multimodal (vision); open weights under Apache 2.0. EU frontier anchor.
• Mistral Magistral Medium Mistral AI · LLM · 2025 — Mistral's first reasoning model, the enterprise version of Magistral (June 2025): transparent step-by-step reasoning that works across languages; 73.6% on AIME 2024 (90% with majority voting @64). Magistral Small is the 24B open-weights version.
• Mistral Medium 3 Mistral AI · LLM · 2025 — Frontier-class multimodal model (May 2025) that Mistral says performs at or above 90% of Claude Sonnet 3.7 across benchmarks at $0.40/M input and $2/M output tokens; strongest in coding and STEM; deployable on any cloud or self-hosted on four or more GPUs.
• Mistral NeMo Mistral AI · LLM · 2024 — Mistral AI's Mistral NeMo (July 18, 2024), a 12B model built with NVIDIA with a 128K-token context and reasoning, world knowledge and coding accuracy Mistral called state of the art for its size; trained for function calling and strong in languages from English and French to Chinese, Arabic and Hindi, it introduced the Tekken tokenizer and quantization-aware training for FP8 inference, and is a drop-in replacement for Mistral 7B. Apache 2.0; retired from Mistral's API July 31, 2026.
• Mistral OCR 3 Mistral AI · Vision/CV · 2025 — Mistral's third-generation document OCR model (mistral-ocr-2512), released Dec 18, 2025; the OCR API added markdown/HTML table formatting, header and footer extraction, and hyperlink output.
• Mistral Saba Mistral AI · LLM · 2025 — 24B regional language model trained on curated data from the Middle East and South Asia; Mistral says it answers more accurately there than models over 5x its size, while being faster and cheaper.
• Mistral Small 3 Mistral AI · LLM · 2025 — 24B open model (Jan 2025) that Mistral says rivals larger models; once quantized it fits a single RTX 4090 or a 32GB MacBook. Apache 2.0.
• Mistral Small 3.1 Mistral AI · LLM · 2025 — 24B open model that adds vision understanding and a 128K context to Mistral Small 3 without losing text performance; can run locally once quantized. Apache 2.0.
• Mistral-7B-Instruct-v0.3 Mistral AI · LLM · 2024 — Mistral AI's Mistral-7B-Instruct-v0.3, an instruction-tuned version of Mistral-7B-v0.3, which differs from v0.2 by extending the vocabulary to 32,768 tokens, supporting the v3 tokenizer and adding function calling; its weights are on Hugging Face under Apache 2.0 for use with mistral-inference or Transformers. Mistral 7B was retired from Mistral's API on March 30, 2025.
• Mistral-7B-v0.2 Mistral AI · LLM · 2024 — Updated Mistral 7B with a 32K context window (up from 8K in v0.1), RoPE theta of 1e6 and no sliding-window attention.
• Mixtral 8x22B Mistral AI · LLM · 2024 — Mistral AI's Mixtral 8x22B (April 17, 2024), a sparse mixture-of-experts model using 39B active parameters out of 141B, fluent in English, French, Italian, German and Spanish, strong at math and coding, natively capable of function calling, with a 64K-token context; Mistral said it was faster than any dense 70B model and more capable than any other open-weight model, and its instruct version scored 90.8% on GSM8K (maj@8). Apache 2.0; retired from Mistral's API March 30, 2025.
• Mixtral 8x7B Mistral AI · LLM · 2023 — Mistral AI's Mixtral 8x7B (December 11, 2023), a sparse mixture-of-experts model whose router picks 2 of 8 expert groups for each token, so it has 46.7B total parameters but uses 12.9B per token; it outperformed Llama 2 70B on most benchmarks with 6x faster inference, matched or beat GPT-3.5 on most standard benchmarks, handled a 32K-token context and English, French, Italian, German and Spanish, and was strong at code. Apache 2.0; retired from Mistral's API March 30, 2025.
• Mixtral 8x7B Instruct Mistral AI · LLM · 2023 — Mistral AI's Mixtral 8x7B Instruct (December 11, 2023), released alongside the Mixtral 8x7B base model and optimized through supervised fine-tuning and direct preference optimization for careful instruction following; it scored 8.30 on MT-Bench, which Mistral called the best open-source result and comparable to GPT-3.5, and it follows any instructions given unless prompted to ban certain outputs. Apache 2.0; Mixtral 8x7B was retired from Mistral's API on March 30, 2025.
• Pixtral Mistral AI · Multimodal LLM · 2024 — Mistral AI's Pixtral 12B (September 17, 2024), a natively multimodal model trained on interleaved image and text data, pairing a new 400M-parameter vision encoder with a 12B decoder based on Mistral NeMo; it accepts variable image sizes and multiple images in a 128K-token context while keeping state-of-the-art text-only performance. Released under Apache 2.0, it is no longer maintained, and Mistral retired it from its API on December 31, 2025.
• Pixtral 12B Mistral AI · Multimodal LLM · 2024 — Mistral AI's natively multimodal Pixtral 12B: a 12B-parameter decoder plus a 400M-parameter vision encoder, trained on interleaved image and text data, with support for variable image sizes and a 128K context.
• Pixtral Large Mistral AI · Multimodal LLM · 2024 — Mistral's 124B open-weights multimodal model (November 2024), built on Mistral Large 2 to add image understanding to its text skills; the Mistral API retired it on May 31, 2026.
• Robostral Navigate Mistral AI · Robotics · 2026 — Mistral's first embodied-navigation model: an 8B robotics model guiding robots through natural-language task instructions using a single RGB camera; claimed SOTA on the R2R-CE benchmark.
• Voxtral Mistral AI · Multimodal LLM · 2025 — Mistral's first audio models (Voxtral Small and Voxtral Mini, Apache 2.0) for speech transcription and spoken-language understanding in chat, plus a Transcription API (Voxtral Mini Transcribe); 32K context.
• ClinicalBERT MIT · Medical NLP · 2019 — ClinicalBERT, publicly released BERT models for clinical text, one for generic clinical notes and one for discharge summaries; its authors found they improved three common clinical NLP tasks over non-specific embeddings but did worse on two de-identification tasks.
• Moises AI Moises · Audio Processing · 2023 — AI music practice. Stem separation.
• Monday.com AI Monday.com · Productivity AI · 2023 — AI automation in Monday.com.
• Kimi Chat Moonshot AI · Chat Interface · 2024 — Kimi, Moonshot AI's AI assistant on the web and in mobile and desktop apps, now powered by Kimi K3, Moonshot's 2.8-trillion-parameter, natively multimodal model with a 1M-token context: beyond chat it runs agent tasks such as deep research, one-click website building, spreadsheets and slide decks, builds playable games, and uses Swarm agent clusters and a Goal mode to work on tasks in parallel; a Kimi Enterprise plan adds team accounts with data-privacy controls.
• Kimi K2 Thinking Moonshot AI · Reasoning LLM · 2025 — Kimi K2 Thinking, Moonshot AI's open-weight (modified MIT) thinking agent built from Kimi K2: a 1T-parameter mixture-of-experts with 32B active and a 256K-token context, trained end to end to interleave step-by-step reasoning with tool calls and to stay on task across 200 to 300 sequential calls, and natively quantized to INT4 for about 2x faster low-latency inference. Moonshot reports 44.9 on HLE (text-only) with tools, 60.2 on BrowseComp and 71.3 on SWE-bench Verified.
• Kimi K2.5 Moonshot AI · Multimodal LLM · 2026 — Kimi K2.5, Moonshot AI's open-weight (modified MIT) native multimodal agentic model, continually pretrained from Kimi-K2-Base on about 15 trillion visual and text tokens: a 1T-parameter mixture-of-experts with 32B active, a 256K-token context and a MoonViT vision encoder, with instant and thinking modes, visual coding and Agent Swarm, which splits tasks across parallel sub-agents. Moonshot reports 50.2 on HLE with tools, 76.8 on SWE-Bench Verified and 78.4 on BrowseComp with Agent Swarm.
• Kimi K2.6 Moonshot AI · Multimodal LLM · 2026 — 1T total params (32B active MoE), 256K context, native multimodal (text/image/video via MoonViT). SWE-bench Verified 80.2% (vendor-reported); native agent-swarm support up to 300 sub-agents. Ranked with DeepSeek V4 Pro among the leading open-weights coding models at its April 2026 launch; Moonshot has since released Kimi K2.7 Code and Kimi K3. Also serves as base model for Cursor Composer 2.5.
• Kimi K2.7 Code Moonshot AI · Code LLM · 2026 — Trillion-parameter open MoE coding model with benchmark jumps over K2.6 and ~30% fewer reasoning tokens; base model for Cognition's SWE-1.7; 289 tok/s on Blackwell NVFP4 serving.
• Kimi K3 Moonshot AI · Multimodal LLM · 2026 — Kimi K3, Moonshot AI's open-weight native multimodal model and, by its account, the world's first open 3T-class model: 2.8T total and 104B active parameters, with 16 of 896 experts used per token, Kimi Delta Attention and Attention Residuals, which Moonshot says give about 2.5x better scaling efficiency than Kimi K2, a 1M-token context and MXFP4 quantization-aware training. Moonshot says it still trails Claude Fable 5 and GPT 5.6 Sol; its API costs $3/$15 per million tokens ($0.30 cached input).
• Moonlight Moonshot AI · LLM · 2025 — Mixture-of-experts model (16B total, 3B active) trained with the Muon optimizer, showing that Muon scales to large-model training. MIT license.
• Moonshot AI Kimi k1.5 Moonshot AI · LLM · 2025 — Moonshot AI's multimodal LLM trained with large-scale reinforcement learning (Jan 2025), with long-context scaling and improved policy optimization as its key training ingredients.
• Moonshot AI Kimi K2 Moonshot AI · LLM · 2025 — Open-weight mixture-of-experts model with 1 trillion total and 32 billion activated parameters, from Moonshot AI; weights and code are released under a Modified MIT License.
• MPT-30B MosaicML · LLM · 2023 — 30B model. Strong code and chat.
• MPT-7B MosaicML · LLM · 2023 — 7B model with a 2,048-token context; the MPT-7B-StoryWriter variant handles 65,536 tokens. Released under Apache 2.0.
• Mubert Mubert · Music Generation · 2023 — AI-generated royalty-free music.
• Murf AI Murf · Speech Synthesis · 2023 — AI voiceover platform. 120+ voices.
• n8n AI n8n · Automation AI · 2023 — Open-source AI workflow automation.
• Napkin AI Napkin AI · Design AI · 2024 — Text-to-diagram and visual generation.
• Dia Nari Labs · Speech Synthesis · 2025 — Nari Labs' Dia, a 1.6B-parameter text-to-speech model that generates realistic dialogue directly from a transcript, with audio conditioning for emotion and tone control and nonverbal sounds such as laughter and coughing; English only, with weights on Hugging Face. Nari Labs has since released Dia2.
• Dia (Nari Labs) Nari Labs · Speech Synthesis · 2025 — Nari Labs' Dia, a 1.6B-parameter text-to-speech model that generates realistic dialogue directly from a transcript, with audio conditioning for emotion and tone control and nonverbal sounds such as laughter and coughing; English only, with weights on Hugging Face. Nari Labs has since released Dia2.
• Donut Naver · Document AI · 2022 — Naver Clova's Donut (Document understanding transformer, ECCV 2022), an OCR-free end-to-end Transformer for visual document understanding, such as document classification and information extraction (document parsing), that needs no off-the-shelf OCR engine; it comes with SynthDoG, a synthetic document generator that makes pretraining flexible across languages and domains.
• Typecast Neosapience · Speech Synthesis · 2023 — AI voice actors for content.
• NightCafe Creator NightCafe · Image Generation · 2021 — AI art platform. Multiple algorithms.
• GPT4All Nomic AI · Local Inference · 2023 — Nomic AI's GPT4All, a desktop application that runs large language models privately on everyday desktops and laptops with no API calls or GPU required; installers cover Windows (including ARM), macOS (best on Apple Silicon) and x86-64 Ubuntu Linux, also on Flathub, and it now supports DeepSeek R1 distillations.
• Nomic Embed Nomic AI · Embeddings · 2024 — Nomic's nomic-embed-text-v1.5 text embedding model, with an 8,192-token context and resizable Matryoshka embeddings (768 dimensions, shrinkable to smaller sizes); prompts need a task prefix such as search_document or search_query, and nomic-embed-vision-v1.5 shares its embedding space.
• Notion AI Notion · Productivity AI · 2023 — AI writing and summarization in Notion.
• Hermes-2-Pro-Mistral-7B Nous Research · LLM · 2024 — Nous Research's Hermes 2 Pro on Mistral 7B, retrained on a cleaned OpenHermes 2.5 dataset plus an in-house function-calling and JSON-mode dataset; Nous reports 90% on a function-calling evaluation built with Fireworks.ai and 84% on structured JSON output. Uses the ChatML prompt format.
• Nous Capybara Nous Research · LLM · 2023 — Nous Research's Nous-Capybara-34B V1.9: Yi-34B with a 200K-token context, fine-tuned for 3 epochs on Capybara, a dataset of about 20K examples of mostly in-house synthesized conversations made with Nous's Amplify-Instruct method; Nous's first 34B and first 200K-context model.
• Nous DeepHermes 3 Preview Nous Research · LLM · 2025 — Nous Research Hermes model (Llama 3.1 8B) that is among the first to unify normal responses and long chain-of-thought reasoning in one model, with improved function calling.
• Nous DisTrO 15B Nous Research · LLM · 2025 — 15B-parameter LLM that Nous Research began pre-training over the internet in December 2024, using its DisTrO and DeMo low-bandwidth optimizers on GPUs at Oracle, Lambda Labs, Northern Data Group, Crusoe Cloud and the Andromeda Cluster, to show that decentralized training works.
• Nous Hermes 4 Nous Research · LLM · 2025 — Hybrid-mode reasoning models from Nous Research built on Llama 3.1 (flagship 405B), post-trained on a new corpus of verified reasoning traces for stronger math, code, STEM and logic.
• OpenHermes 2.5 Nous Research · LLM · 2023 — OpenHermes 2.5, a Mistral 7B fine-tune trained on 1,000,000 entries of mainly GPT-4-generated data plus filtered open datasets; added code instructions (an estimated 7% to 14% of the data) raised its HumanEval pass@1 from 43% (OpenHermes 2) to 50.7% and boosted several non-code benchmarks. Uses the ChatML prompt format.
• NovelAI NovelAI · AI Writing · 2021 — AI storytelling and image generation.
• Cosmos 1.0 NVIDIA · Video Generation · 2025 — NVIDIA's first Cosmos world foundation models, launched at CES 2025: open models that generate physics-based video from text, image or video inputs, as well as robot sensor or motion data, to help developers train and test robots and autonomous vehicles.
• Cosmos Nemotron 34B NVIDIA · Video Generation · 2025 — Vision language model in NVIDIA's Cosmos Nemotron family, announced at CES 2025, that lets AI agents analyze and respond to images and video; offered as an NVIDIA NIM microservice.
• Instant-NGP NVIDIA · 3D Rendering · 2022 — NVIDIA's Instant Neural Graphics Primitives (instant-ngp, SIGGRAPH 2022), which trains and renders four neural graphics primitives (neural radiance fields, signed distance functions, neural images and neural volumes) as small MLPs with a multiresolution hash input encoding on the tiny-cuda-nn framework, fast enough to train a NeRF of a fox in under 5 seconds.
• Llama-3.1-Nemotron-Ultra-253B NVIDIA · LLM · 2025 — 253B, NAS-distilled from Llama 3.1 405B. Reasoning on/off toggle. Open weights. Single-node inference focus.
• Llama-3.3-Nemotron-Super-49B-v1 NVIDIA · LLM · 2025 — 49B, NAS-distilled from Llama 3.3 70B. Reasoning toggle. Optimized for single-GPU deployment. Open weights.
• Nemotron 3 Nano NVIDIA · LLM · 2025 — Nemotron 3 Nano 30B-A3B: hybrid Mamba-2/Transformer MoE with 30B total and ~3.5B active parameters, up to 1M-token context, trained on 25T tokens; released Dec 15, 2025 with open weights, training data and recipes, and a reasoning on/off switch. Commercial use permitted.
• Nemotron 3 Super NVIDIA · LLM · 2026 — 120B-total / 12B-active hybrid Mamba-Transformer latent-MoE with multi-token prediction, built for multi-agent reasoning, coding and long-context work; pretrained on 25T seen tokens; open weights on Hugging Face.
• Nemotron 3 Ultra NVIDIA · LLM · 2026 — 550B-parameter sparse MoE with 55B active, built for long-running agentic harnesses (OpenCode, Hermes, OpenClaw): hybrid Mamba/Transformer architecture and an unusually complete open release — weights, training data, recipes, reward model, quantized checkpoint.
• Nemotron 3.5 ASR NVIDIA · Speech Recognition · 2026 — 600M-parameter open multilingual streaming speech-to-text for voice agents: 40 languages, ~17x throughput vs Parakeet-style baselines at half the size.
• Nemotron Nano 8B NVIDIA · LLM · 2025 — NVIDIA's Llama-3.1-Nemotron-Nano-8B-v1, a reasoning model derived from Llama 3.1 8B Instruct and post-trained for reasoning, chat, RAG and tool calling (supervised fine-tuning plus RLOO and RPO reinforcement learning); 128K context, fits on a single RTX GPU and is ready for commercial use.
• Nemotron-4 NVIDIA · LLM · 2024 — NVIDIA's 340B-parameter model (base and instruct versions) built for synthetic data generation pipelines that create training data for other LLMs; 4K-token context.
• Obviously AI Obviously AI · ML Platform · 2023 — No-code AI/ML predictions. Obviously AI has since become Zams; obviously.ai now archives the service as it ran through 2025.
• Ollama Ollama · Local Inference · 2023 — Run LLMs locally. Simple CLI.
• ChatGPT OpenAI · Chat Interface · 2022 — OpenAI's ChatGPT, launched November 30, 2022 as a free research preview: a sibling of InstructGPT, fine-tuned from the GPT-3.5 series with reinforcement learning from human feedback, whose dialogue format let it answer follow-up questions, admit its mistakes, challenge incorrect premises and reject inappropriate requests. It now has paid plans and an ad-supported free tier, which OpenAI said on August 31, 2026 helps keep ChatGPT available to more than 1 billion weekly active users.
• ChatGPT Agent OpenAI · Agent · 2025 — OpenAI's ChatGPT agent (July 17, 2025), an agent mode that combined Operator's ability to act on websites with deep research's skill at synthesizing information, carrying out multi-step tasks on its own virtual computer; launched for Pro (400 messages a month), Plus and Team users (40), it set state-of-the-art scores of 41.6% on Humanity's Last Exam and 68.9% on BrowseComp, and absorbed Operator. OpenAI's help center now says ChatGPT agent is no longer available and points users to ChatGPT Work.
• ChatGPT Canvas OpenAI · AI Writing/Coding · 2024 — ChatGPT canvas (October 3, 2024), a GPT-4o-based interface opening writing and coding projects in a separate window, with shortcuts to suggest edits, adjust length or reading level, add final polish or emojis, review code, add logs or comments, fix bugs and port code to another language; it launched in beta for Plus and Team users. Since May 28, 2026 newer models handle writing and code in chat responses instead, and paid users keep canvas only through legacy models until those are sunset.
• ChatGPT Images 2.0 OpenAI · Image Generation · 2026 — Next-generation in-ChatGPT image generation. Higher fidelity, better instruction following, improved typography and multi-object composition. Announced Apr 21, 2026.
• CLIP OpenAI · Vision/CV · 2021 — OpenAI's CLIP (January 5, 2021), a neural network that learns visual concepts from natural-language supervision on image-text pairs from the internet and can be instructed in natural language to pick the most relevant text for an image, a zero-shot ability like GPT-2's and GPT-3's; it matched the original ResNet-50 on ImageNet without any of its 1.28M labeled examples, but struggled with abstract tasks such as counting objects. Code and models are on GitHub under the MIT license.
• CLIP (ViT B/32) OpenAI · Vision/CV · 2021 — OpenAI's CLIP with a ViT-B/32 image encoder: CLIP learns image representations by predicting which caption goes with which image across 400 million internet image-text pairs, so natural-language prompts enable zero-shot classification; OpenAI reports zero-shot ImageNet accuracy matching the original ResNet-50 without its 1.28 million labeled examples, and released the code and weights.
• CLIP (ViT L/14) OpenAI · Vision/CV · 2021 — CLIP ViT-L/14, a Vision Transformer version of OpenAI's CLIP (Contrastive Language-Image Pre-Training) from January 2021: an image encoder and a text Transformer trained contrastively on 400 million image-text pairs from the internet to match images with captions, so it classifies images zero-shot from natural-language labels. OpenAI found CLIP matched the original ResNet-50 on ImageNet zero-shot without using any of its 1.28 million labeled examples; code and weights are MIT-licensed.
• Codex (original) OpenAI · Code LLM · 2021 — OpenAI's original Codex, a GPT language model fine-tuned on public GitHub code; a distinct production version powered GitHub Copilot. On OpenAI's HumanEval benchmark it solved 28.8% of problems in one attempt (GPT-3 solved 0%) and 70.2% with 100 samples per problem, and OpenAI noted weaknesses with long chains of operations and binding operations to variables.
• codex-1 OpenAI · Code LLM · 2025 — OpenAI's version of o3 optimized for software engineering, trained with reinforcement learning on real-world coding tasks; it powered the cloud-based Codex agent in ChatGPT at launch in May 2025.
• DALL-E 3 OpenAI · Image Generation · 2023 — OpenAI's DALL·E 3, its 2023 image model, which understood significantly more nuance and detail than its earlier systems and was built natively on ChatGPT, which could write detailed prompts for it; OpenAI said the images people create with it are theirs to use, and the model was designed to decline requests for public figures by name or images in a living artist's style, with an opt-out for creators. Offered in ChatGPT and the API, it was removed from the OpenAI API on May 12, 2026.
• DALL·E OpenAI · Image Generation · 2021 — OpenAI's original DALL·E, which generates images from text with a transformer that autoregressively models text and image tokens as a single stream of data; with enough data and scale, OpenAI found it competitive with earlier domain-specific models in zero-shot evaluation.
• DALL·E 2 OpenAI · Image Generation · 2022 — OpenAI's DALL·E 2 (2022), an AI system that creates realistic images and art from a natural-language description, combining concepts, attributes and styles at 4x the resolution of the original DALL·E, and able to edit images (inpainting), extend them beyond the frame (outpainting) and make variations; OpenAI limited its ability to generate violent, hate or adult images and photorealistic faces of real people. It opened in beta in July 2022 and was removed from the OpenAI API on May 12, 2026.
• GPT Image 1 OpenAI · Image Generation · 2025 — OpenAI's GPT Image 1, released in the API on April 23, 2025 as the natively multimodal model behind the image generation OpenAI had added to ChatGPT a month earlier, which over 130 million users used to create more than 700 million images in its first week; it follows custom guidelines, draws on world knowledge and renders text accurately, cost about $0.02-$0.19 per square image, and was adopted by Adobe, Figma and others. Now superseded, it leaves the API October 23, 2026.
• GPT Image 2.5 Flare OpenAI · Image Generation · 2026 — GPT Image 2.5 Flare (September 8, 2026), the default, fastest model in OpenAI's GPT Image 2.5 family, launched with ChatGPT Images 2.5: it makes higher-quality images than GPT Image 2 at 50% lower latency, edits only what is asked, keeps changes consistent across multi-turn edits and handles transparent backgrounds. It takes text and image input, offers quality settings from low to max, and costs $5 per million text-input and $30 per million image-output tokens.
• GPT Image 2.5 Sunburst OpenAI · Image Generation · 2026 — GPT Image 2.5 Sunburst (September 8, 2026), OpenAI's most capable image generation and editing model and the precision option in the GPT Image 2.5 family, launched with ChatGPT Images 2.5: aimed at premium work like campaign creative and product imagery, it trades longer generation times for extra edit precision, keeping unrequested areas, earlier edits and reference subjects intact. It takes text and image input and costs $5 per million text-input and $30 per million image-output tokens.
• GPT-1 OpenAI · LLM · 2018 — GPT-1, the model from OpenAI's paper Improving Language Understanding by Generative Pre-Training: a 12-layer decoder-only Transformer (768-dimensional states, 12 attention heads) pre-trained as a language model on BooksCorpus, over 7,000 unpublished books, then fine-tuned for each task; it beat task-specific models and set the state of the art on 9 of 12 language understanding tasks, including gains of 8.9% on the Stories Cloze Test and 5.7% on RACE. OpenAI released its code and model.
• GPT-2 OpenAI · LLM · 2019 — 1.5B parameters. OpenAI initially withheld it over misuse fears — the first time an AI lab self-censored a release. Sparked the "too dangerous to publish" debate.
• GPT-2 (1.5B) OpenAI · LLM · 2019 — OpenAI's GPT-2 (February 14, 2019), a 1.5-billion-parameter language model trained to predict the next word in 40GB of internet text from 8 million web pages, which generated coherent paragraphs, set state-of-the-art results on many language modeling benchmarks and did rudimentary reading comprehension, translation, question answering and summarization without task-specific training; citing misuse concerns, OpenAI released it in stages, publishing the full model and code on November 5, 2019.
• GPT-3 OpenAI · LLM · 2020 — GPT-3, OpenAI's 175-billion-parameter autoregressive language model, 10 times larger than any previous non-sparse model: its 2020 paper showed it could do many tasks few-shot, from a handful of examples in the prompt and without gradient updates or fine-tuning, with strong results on translation, question answering and cloze tasks and news articles that human evaluators had difficulty telling from human writing. OpenAI shut down its original API models, including davinci, on January 4, 2024.
• GPT-3 (ada) OpenAI · LLM · 2020 — OpenAI's ada, the lowest-priced of the original GPT-3 base models in OpenAI's API ($0.40 per 1M tokens, versus $0.50 for babbage, $2.00 for curie and $20.00 for davinci); OpenAI shut it down on Jan 4, 2024 and recommends babbage-002 as the replacement.
• GPT-3 (babbage) OpenAI · LLM · 2020 — OpenAI's babbage, one of the original GPT-3 base models in OpenAI's API, priced at $0.50 per 1M tokens (ada $0.40, curie $2.00, davinci $20.00); OpenAI shut it down on Jan 4, 2024 and recommends babbage-002 as the replacement.
• GPT-3 (curie) OpenAI · LLM · 2020 — OpenAI's curie, one of the original GPT-3 base models in OpenAI's API, priced at $2.00 per 1M tokens, between babbage ($0.50) and davinci ($20.00); OpenAI shut it down on Jan 4, 2024 and recommends davinci-002 as the replacement.
• GPT-3 175B (davinci) OpenAI · LLM · 2020 — OpenAI's GPT-3, an autoregressive language model with 175 billion parameters, 10x more than any previous non-sparse language model, introduced in a May 2020 paper showing that scaling greatly improves few-shot performance: given a few examples and no fine-tuning, it handled translation, question answering, cloze tasks and on-the-fly reasoning, and wrote news articles evaluators struggled to tell from human ones. OpenAI's API, launched in private beta on June 11, 2020, ran GPT-3-family models.
• GPT-3.5 Turbo OpenAI · LLM · 2022 — OpenAI's GPT-3.5 Turbo, the model family behind ChatGPT, opened to developers on March 1, 2023 as gpt-3.5-turbo at $0.002 per 1K tokens, 10x cheaper than OpenAI's existing GPT-3.5 models, after a 90% cut in ChatGPT's costs since December 2022; optimized for chat but also good at non-chat tasks, its current snapshot has a 16,385-token context at $0.50/$1.50 per million tokens. OpenAI has recommended GPT-4o mini instead since July 2024 and removes it from the API October 23, 2026.
• GPT-4 OpenAI · Multimodal LLM · 2023 — OpenAI's GPT-4 (March 14, 2023), a large multimodal model accepting image and text inputs and producing text, which showed human-level performance on many professional and academic benchmarks, passing a simulated bar exam around the top 10% of test takers where GPT-3.5 scored around the bottom 10%, after 6 months of iterative alignment; it reached ChatGPT Plus and an API waitlist in 8,192- and 32,768-token versions, with image input still a research preview. It leaves the API October 23, 2026.
• GPT-4 Turbo OpenAI · Multimodal LLM · 2023 — OpenAI's GPT-4 Turbo, announced at its first DevDay on November 6, 2023 as gpt-4-1106-preview: a more capable GPT-4 with knowledge of world events up to April 2023 and a 128K-token context (more than 300 pages of text), priced 3x lower for input and 2x lower for output than GPT-4 at $0.01/$0.03 per 1K tokens, with JSON mode and parallel function calling; its gpt-4-turbo-2024-04-09 snapshot also takes images. OpenAI removes it from the API October 23, 2026.
• GPT-4 Vision OpenAI · Multimodal LLM · 2023 — GPT-4 with vision (GPT-4V), which lets users instruct GPT-4 to analyze images they provide; OpenAI published its system card on September 25, 2023 as image input began rolling out to ChatGPT Plus and Enterprise users, powered by multimodal GPT-3.5 and GPT-4 and informed by work with the Be My Eyes app for blind and low-vision people, and opened it to developers as gpt-4-vision-preview at DevDay in November 2023. That preview API was shut down December 6, 2024.
• GPT-4.1 OpenAI · Multimodal LLM · 2025 — OpenAI's GPT-4.1 (April 14, 2025), the flagship of a family launched only in the API alongside GPT-4.1 mini and nano, which outperformed GPT-4o across the board with major gains in coding and instruction following: 54.6% on SWE-bench Verified (21.4 points above GPT-4o) and 38.3% on Scale's MultiChallenge, with a 1M-token context and a June 2024 knowledge cutoff, at $2/$8 per million tokens. OpenAI removes it from the API January 14, 2027, recommending GPT-6 Luna.
• GPT-4.1 mini OpenAI · LLM · 2025 — OpenAI's GPT-4.1 mini (April 14, 2025), the middle model of the GPT-4.1 family launched in the API, which OpenAI said matched or beat GPT-4o in intelligence evaluations while cutting latency nearly in half and cost by 83%; it excels at instruction following and tool calling without a reasoning step, takes text and images with a context window of about 1M tokens (32,768 output) and costs $0.40/$1.60 per million tokens. OpenAI removes it from the API January 14, 2027, recommending GPT-6 Luna.
• GPT-4.1 nano OpenAI · LLM · 2025 — OpenAI's GPT-4.1 nano (April 14, 2025), the smallest GPT-4.1 model and, at launch, OpenAI's fastest and cheapest, for tasks that demand low latency; it scored 80.1% on MMLU, 50.3% on GPQA and 9.8% on Aider polyglot coding, excels at instruction following and tool calling, takes text and images with a context window of about 1M tokens, and costs $0.10/$0.40 per million tokens. OpenAI removes it from the API October 23, 2026, recommending GPT-5.6 Luna.
• GPT-4.5 OpenAI · Multimodal LLM · 2025 — OpenAI's GPT-4.5 (February 27, 2025), a research preview it called its largest and best model for chat yet, built by scaling unsupervised learning for broader knowledge, a deeper understanding of the world, greater 'EQ' and fewer hallucinations; it reached ChatGPT Pro users first, then Plus, Team, Enterprise and Edu, and API developers on all paid tiers at $75/$150 per million tokens. The gpt-4.5-preview API was shut down July 14, 2025 in favor of GPT-4.1.
• GPT-4b micro OpenAI · Science/Biology · 2025 — Specialized protein-engineering model developed with Retro Biosciences. Initialized from GPT-4o, fine-tuned on protein data. Not available via public API or ChatGPT. Non-public, biology-specific — not a general-purpose LLM.
• GPT-4o OpenAI · Multimodal LLM · 2024 — OpenAI's GPT-4o (May 13, 2024), its 'omni' model accepting any combination of text, audio, image and video and generating text, audio and images, answering audio in as little as 232 milliseconds (320 on average); it matched GPT-4 Turbo on English text and code while being twice as fast and half the price in the API, and came to ChatGPT's free tier. The API model takes text and images with a 128K-token context; OpenAI removes its last snapshots on January 14, 2027.
• GPT-4o mini OpenAI · LLM · 2024 — OpenAI's GPT-4o mini (July 18, 2024), its most cost-efficient small model at launch: priced at $0.15/$0.60 per million tokens, more than 60% below GPT-3.5 Turbo, which it replaced for ChatGPT Free, Plus and Team users, it scored 82.0% on MMLU, 87.0% on MGSM and 87.2% on HumanEval, takes text and images with a 128K-token context and 16K output, and was the first model to apply OpenAI's instruction hierarchy against jailbreaks. It leaves the API January 14, 2027.
• GPT-5 OpenAI · Multimodal LLM · 2025 — OpenAI's GPT-5, launched August 2025 as its model for coding, reasoning and agentic tasks, with configurable reasoning effort (minimal to high), a 400,000-token context window and 128,000 max output tokens; now deprecated in the API, where OpenAI recommends GPT-6 Astra.
• GPT-5 mini OpenAI · LLM · 2025 — Faster, cost-efficient GPT-5 variant for well-defined tasks: 400K context window, 128K max output, text and image input, reasoning-token support; May 31, 2024 knowledge cutoff.
• GPT-5 nano OpenAI · LLM · 2025 — Fastest, cheapest GPT-5 variant, aimed at summarization and classification: 400K context window, 128K max output, text and image input; May 31, 2024 knowledge cutoff. OpenAI now points new speed- and cost-sensitive workloads to newer small models.
• GPT-5.1 OpenAI · Multimodal LLM · 2025 — OpenAI's GPT-5.1 (November 12, 2025), an update to GPT-5 that came to ChatGPT as GPT-5.1 Instant, warmer, smarter and better at following instructions, able to decide when to think before responding, and GPT-5.1 Thinking, whose answers are easier to understand, alongside new personality presets such as Professional, Candid and Quirky; in the API it is a model for coding and agentic tasks with a 400K-token context at $1.25/$10 per million tokens. OpenAI removes it from the API April 1, 2027.
• GPT-5.2 OpenAI · Multimodal LLM · 2025 — OpenAI's GPT-5.2 (December 11, 2025), billed as its most advanced frontier model for professional work and long-running agents and released in ChatGPT as GPT-5.2 Instant, Thinking and Pro for paid plans; GPT-5.2 Thinking scored 70.9% on GDPval, 80.0% on SWE-bench Verified and 92.4% on GPQA Diamond, with responses containing errors 30% less common than GPT-5.1 Thinking's. In the API it has a 400K-token context at $1.75/$14 per million tokens, and OpenAI now calls it its previous flagship.
• GPT-5.3-Codex OpenAI · Code LLM · 2026 — Coding/agentic specialist positioned against Claude Opus 4.6; ~25% faster than Opus 4.6 per OpenAI. First OpenAI model "instrumental in creating itself." Codex-Spark mini variant Feb 12.
• GPT-5.4 OpenAI · Multimodal LLM · 2026 — OpenAI's GPT-5.4 (March 5, 2026), billed as its most capable and efficient frontier model for professional work and its first general-purpose model with native computer use, scoring 75.0% on OSWorld-Verified against GPT-5.2's 47.3% and 83.0% on GDPval; its individual claims are 33% less likely to be false than GPT-5.2's. It supports up to 1M tokens of context in the API and Codex at $2.50/$15 per million tokens, and came to ChatGPT as GPT-5.4 Thinking alongside GPT-5.4 Pro.
• GPT-5.4 mini OpenAI · LLM · 2026 — Small GPT-5.4 model optimized for coding, computer use, multimodal understanding and sub-agents; OpenAI says it is more than 2x faster than GPT-5 mini and approaches GPT-5.4 on SWE-Bench Pro and OSWorld-Verified. 400K context, 128K max output, Aug 31, 2025 knowledge cutoff; ChatGPT fallback when GPT-5.4 limits are reached.
• GPT-5.4 nano OpenAI · LLM · 2026 — Smallest, fastest GPT-5.4 model for classification, data extraction, ranking and simple sub-agent tasks; API-only (not in ChatGPT). 400K context, 128K max output, Aug 31, 2025 knowledge cutoff.
• GPT-5.5 OpenAI · Multimodal LLM · 2026 — Codename "Spud." Launched April 2026 as OpenAI's flagship; OpenAI's model guide now recommends the GPT-6 family. 82.7% Terminal-Bench 2.0, 51.7% FrontierMath T1-3; narrowly beat Claude Mythos Preview on Terminal-Bench at launch.
• GPT-5.5 Pro OpenAI · Multimodal LLM · 2026 — OpenAI's GPT-5.5 Pro (April 23, 2026), the version of GPT-5.5 that uses more compute to think harder, for harder questions and higher-accuracy work, rolled out to ChatGPT Pro, Business and Enterprise users; testers found its answers more comprehensive, well-structured and accurate than GPT-5.4 Pro's, especially in business, legal, education and data science, and it scored 90.1% on BrowseComp. In the API it runs through the Responses API with a 1.05M-token context at $30/$180 per million tokens.
• GPT-5.6 Luna OpenAI · LLM · 2026 — OpenAI's GPT-5.6 Luna (July 9, 2026), the fastest and most affordable of the GPT-5.6 models alongside Sol and Terra, roughly the nano tier of earlier GPT-5 families; OpenAI says it nearly matches GPT-5.5's peak performance at less than half the estimated cost. It takes text and images with a 1.05M-token context and 128K-token output at $0.20/$1.20 per million tokens after an 80% price cut, and became the default model for ChatGPT Free and Go users in August 2026.
• GPT-5.6 Sol OpenAI · LLM · 2026 — GPT-5.6 Sol (July 9, 2026), the flagship of OpenAI's GPT-5.6 family alongside the cheaper Terra and the faster Luna, built for complex professional work: a 1.05M-token context, 128K output tokens, reasoning effort up to max and an ultra mode that runs four agents in parallel. OpenAI reports 64.6% on SWE-Bench Pro, 88.8% on Terminal-Bench 2.1 and 7.78% on ARC-AGI-3; it also runs on Cerebras at up to 750 tokens a second, and OpenAI says Sol's own kernel work cut its serving cost by 20%.
• GPT-5.6 Terra OpenAI · LLM · 2026 — OpenAI's GPT-5.6 Terra (July 9, 2026), the balanced middle tier of the GPT-5.6 family between Sol and Luna, roughly the mini tier of earlier GPT-5 families, which OpenAI says is competitive with GPT-5.5 at lower cost, scoring 63.4% on SWE-Bench Pro and 92.9% on GPQA Diamond; it takes text and images with a 1.05M-token context and 128K-token output, costs $2/$12 per million tokens after a 20% price cut, and was the GPT-5.6 model for ChatGPT Free and Go users at launch.
• GPT-6 Astra OpenAI · LLM · 2026 — OpenAI's next-generation frontier flagship and first of the GPT-6 family: 1,050,000-token context, 128K max output. 58.18% on Terminal-Bench 4.0; a restricted cyber tier posted a perfect ExploitBench score.
• GPT-6 Luna OpenAI · LLM · 2026 — Lowest-cost GPT-6 tier for high-volume, clearly scoped tasks. OpenAI reports DeepSWE v1.1 66.6% at max effort and says that at higher effort it matches GPT-5.6 Sol on its factuality eval at about a hundredth of the cost.
• GPT-6 Sol OpenAI · LLM · 2026 — Mid-tier GPT-6 model trained with methods similar to GPT-6 Astra. OpenAI reports about half as many factual mistakes as GPT-5.6 Sol on its internal factuality eval; AutomationBench 33.2% (xhigh), Agents' Last Exam 56.4% (max), DeepSWE v1.1 68.8% (max), OSWorld 2.0 offline 60.5% (xhigh). Improved prompt caching for agents.
• GPT-Live OpenAI · Speech AI · 2026 — GPT-Live (July 8, 2026), OpenAI's full-duplex voice models for ChatGPT Voice: they listen while speaking and decide many times a second whether to talk, keep listening, pause, interrupt or call a tool, handing web searches and harder reasoning to a frontier model (GPT-5.5 at launch). GPT-Live-1 is the default for Go, Plus and Pro users and GPT-Live-1 mini for Free users, with nine remastered voices and live translation; GPT-Live-1 joined the API on September 10, 2026 ($0.05 a minute).
• GPT-Live-Transcribe OpenAI · Speech Recognition · 2026 — OpenAI's GPT-Live-Transcribe (released July 28, 2026 alongside GPT-Transcribe), a streaming speech-to-text model for apps that need low-latency transcript deltas from live audio, with tunable latency, free-form context, keyword hints for names and domain terms, and multiple expected input languages; it costs $0.017 per minute, and OpenAI recommends it or GPT-Transcribe in place of whisper-1 and its gpt-4o transcription models, which shut down February 26, 2027.
• GPT-OSS 120B OpenAI · LLM · 2025 — OpenAI's gpt-oss-120b (August 5, 2025), the larger of its two Apache 2.0 open-weight gpt-oss reasoning models: a mixture-of-experts model with 117B parameters (5.1B active per token) that OpenAI says reaches near-parity with o4-mini on core reasoning benchmarks while running on a single 80 GB GPU such as an H100. It has a 128K-token context, adjustable reasoning effort, full chain-of-thought access, tool use such as web search and Python execution, and fine-tuning, with weights on Hugging Face.
• GPT-OSS 20B OpenAI · LLM · 2025 — OpenAI's gpt-oss-20b (August 5, 2025), the smaller of its two Apache 2.0 open-weight gpt-oss reasoning models, built for lower-latency, local or specialized use: a mixture-of-experts model with 21B parameters (3.6B active per token) that OpenAI says delivers results similar to o3-mini on common benchmarks and runs within 16 GB of memory. Like gpt-oss-120b it has a 128K-token context, adjustable reasoning effort, full chain-of-thought access, tool use and fine-tuning.
• GPT-Rosalind OpenAI · Science/Biology · 2026 — Frontier reasoning model purpose-built for life sciences, biology, drug discovery, and translational medicine. Optimized for chemistry, protein engineering, and genomics workflows with enhanced tool use.
• GPT-Transcribe OpenAI · Speech Recognition · 2026 — Context-aware batch ASR replacing the 4o-era models: 8.98% word error rate vs Whisper-1's 15.21% (-41%); multilingual errors roughly halved across 22 languages. Context prompting (keywords, language hints, prior turns) measurably lifts semantic accuracy on names, numbers and technical terms in noisy audio.
• InstructGPT 175B OpenAI · LLM · 2022 — OpenAI's InstructGPT models, announced January 27, 2022 and described in a March 2022 paper: GPT-3 fine-tuned with humans in the loop, first on labeler demonstrations and then with reinforcement learning from human feedback on ranked outputs, to follow user intentions far better while being more truthful and less toxic; labelers preferred the 1.3B InstructGPT model's outputs to 175B GPT-3's despite 100x fewer parameters. They became the API's default models, though they still made up facts.
• Jukebox OpenAI · Music Generation · 2020 — OpenAI's Jukebox, which generates music with singing as raw audio: a multi-scale VQ-VAE compresses audio into discrete codes that autoregressive Transformers model, producing coherent songs up to several minutes long that can be steered by artist, genre and unaligned lyrics; OpenAI released 5B and 1B-parameter models, code and thousands of samples.
• OpenAI gpt-oss-120b OpenAI · LLM · 2025 — OpenAI's larger open-weight reasoning model: 117B parameters (5.1B active) that fits on a single 80GB GPU. Apache 2.0; built for reasoning, agentic tasks and tool use.
• OpenAI gpt-oss-20b OpenAI · LLM · 2025 — OpenAI gpt-oss-20b, the smaller of OpenAI's two open-weight gpt-oss reasoning models, built for low-latency and local use: 21B parameters with 3.6B active, MXFP4-quantized mixture-of-experts weights that run within 16GB of memory, a 131K-token context, configurable low, medium or high reasoning effort with full chain-of-thought access, and native function calling, web browsing, Python execution and structured outputs. It uses OpenAI's harmony format; Apache 2.0, fine-tunable.
• OpenAI o1 OpenAI · Reasoning LLM · 2024 — OpenAI o1, an o-series reasoning model trained with reinforcement learning to work through a long internal chain of thought before answering. OpenAI unveiled it on September 12, 2024, releasing o1-preview in ChatGPT and to trusted API users, and reported that o1 scored 83% on AIME 2024 with 64-sample consensus (GPT-4o: 12%), reached the 89th percentile on Codeforces and beat human PhD experts on GPQA Diamond. Its API version ($15/$60 per million tokens, 200K context) shuts down October 23, 2026.
• OpenAI o1-2024-12-17 OpenAI · Reasoning LLM · 2024 — o1-2024-12-17, the snapshot of OpenAI o1 released to the API in December 2024, a post-trained version of the o1 that reached ChatGPT two weeks earlier: it added function calling, structured outputs, developer messages, image input and a reasoning_effort setting, used 60% fewer reasoning tokens than o1-preview on average, and scored 79.2% on AIME 2024 (o1-preview: 42.0%) and 48.9% on SWE-bench Verified. It rolled out first to usage tier 5 and shuts down October 23, 2026.
• OpenAI o1-mini OpenAI · Reasoning LLM · 2024 — OpenAI's o1-mini (September 12, 2024), a cost-efficient reasoning model for STEM, especially math and coding, that came close to o1 on AIME (70.0% against 74.4%) and reached a 1650 Codeforces Elo, at 80% less than o1-preview's cost, though its non-STEM factual knowledge was comparable to small LLMs; it reached ChatGPT Plus, Team, Enterprise and Edu users and tier-5 API developers with a 128K-token context and 65,536-token output. It was removed from the API October 27, 2025.
• OpenAI o1-preview OpenAI · Reasoning LLM · 2024 — OpenAI's o1-preview (September 12, 2024), the research preview of its first o-series reasoning models, trained with reinforcement learning to think before answering by producing a long internal chain of thought; it scored 84/100 on one of OpenAI's hardest jailbreaking tests against GPT-4o's 22. It launched for ChatGPT Plus and Team users and tier-5 API developers at $15/$60 per million tokens with a 128K-token context, without web browsing, and was removed from the API July 28, 2025.
• OpenAI o3 OpenAI · Reasoning LLM · 2025 — OpenAI reasoning model for complex tasks, which OpenAI describes as well-rounded across domains and setting a new standard for math, science, coding and visual reasoning; also strong at technical writing and instruction following. 200K context window, 100K max output; succeeded by GPT-5.
• OpenAI o3-mini OpenAI · Reasoning LLM · 2025 — Small reasoning model offering high intelligence at the same cost and latency targets as o1-mini, with Structured Outputs, function calling and Batch API support; 200K context window.
• OpenAI o3-pro OpenAI · Reasoning LLM · 2025 — OpenAI's o3-pro (June 10, 2025), a version of its o3 reasoning model that uses more compute to think longer and give the most reliable responses; it replaced o1-pro for ChatGPT Pro and Team users, with Enterprise and Edu a week later, and expert reviewers preferred it to o3 in every tested category. In the API it runs only through the Responses API, with a 200K-token context and 100K output, at $20/$80 per million tokens; OpenAI shuts it down December 11, 2026, recommending GPT-5.6 Sol.
• OpenAI o4-mini OpenAI · Reasoning LLM · 2025 — Small o-series reasoning model optimized for fast, cost-efficient reasoning, with especially efficient performance in coding and visual tasks; 200K context window. Succeeded by GPT-5 mini.
• OpenAI text-embedding-3-large OpenAI · Embeddings · 2024 — OpenAI's text-embedding-3-large (January 25, 2024), its most capable embedding model for English and non-English text, producing embeddings of up to 3,072 dimensions; over text-embedding-ada-002 it raised average scores from 31.4% to 54.9% on the MIRACL multilingual retrieval benchmark and from 61.0% to 64.6% on MTEB, and its embeddings can be shortened with a dimensions parameter, still beating a full 1,536-dimension ada-002 embedding on MTEB at 256 dimensions. $0.13 per million tokens.
• OpenAI text-embedding-3-small OpenAI · Embeddings · 2024 — OpenAI's text-embedding-3-small (January 25, 2024), a highly efficient embedding model that upgraded text-embedding-ada-002, raising average scores from 31.4% to 44.0% on the MIRACL multilingual retrieval benchmark and from 61.0% to 62.3% on MTEB while cutting the price 5x to $0.02 per million tokens; it produces 1,536-dimension embeddings by default that can be shortened with a dimensions parameter, for search, clustering, recommendations, anomaly detection and classification.
• OpenAI text-embedding-ada-002 OpenAI · Embeddings · 2022 — OpenAI's text-embedding-ada-002 (December 15, 2022), which replaced five separate models for text search, text similarity and code search and outperformed OpenAI's previous most capable embedding model, Davinci, at most tasks at a 99.8% lower price; it raised context length fourfold to 8,192 tokens and produced 1,536-dimension embeddings, one-eighth the size of davinci-001's. OpenAI recommended text-embedding-3-small from January 2024 but did not deprecate it; it costs $0.10 per million tokens.
• Point-E OpenAI · 3D Generation · 2022 — OpenAI's Point-E, a system that generates 3D point clouds from prompts: its main models create point clouds from synthetic view images, a small, lower-quality model works directly from text, and an SDF regression model turns point clouds into meshes.
• Shap-E OpenAI · 3D Generation · 2023 — OpenAI's Shap-E, which generates conditional 3D implicit functions, producing 3D models from a text prompt or from a synthetic view image; OpenAI released the code and model weights.
• Sora OpenAI · Video Generation · 2024 — Text-to-video generation. Up to 1 min video. Storyboard mode. In-video editing. RETIRED: Sora app/web discontinued Apr 26, 2026; API access ends Sep 24, 2026 (OpenAI, announced Mar 2026).
• Sora 2 OpenAI · Video Generation · 2025 — Synchronized dialogue and sound effects. Physics simulation improvements. RETIRED: Sora app/web discontinued Apr 26, 2026; API access ends Sep 24, 2026 (OpenAI, announced Mar 2026).
• Whisper OpenAI · Speech Recognition · 2022 — OpenAI's Whisper (September 21, 2022), an open-source speech recognition model trained on 680,000 hours of multilingual web audio, about a third non-English, that approaches human-level robustness and accuracy on English; it handles language identification, phrase-level timestamps, multilingual transcription and translation into English, making 50% fewer errors than LibriSpeech-specialized models across diverse datasets. The whisper-1 API ($0.006 a minute) closes February 26, 2027.
• Whisper Large v3 OpenAI · Speech Recognition · 2023 — OpenAI's Whisper large-v3 speech recognition and translation model: the large-v2 architecture with 128 Mel frequency bins and a new Cantonese language token, trained on 1 million hours of weakly labeled and 4 million hours of pseudo-labeled audio; the card reports 10% to 20% fewer errors than large-v2 across many languages.
• Whisper Large v3 Turbo OpenAI · Speech Recognition · 2024 — OpenAI's Whisper large-v3-turbo speech recognition and translation model: a pruned and fine-tuned Whisper large-v3 with 4 decoder layers instead of 32, which makes it much faster at the cost of a minor drop in quality.
• MOSS-VL-Realtime OpenMOSS · Multimodal LLM · 2026 — Open-source 11B vision-language model (~22.7GB) built for real-time streaming video that proactively speaks up or deliberately stays silent instead of only answering prompts; reported SOTA on all three open proactivity benchmarks; base model included in the release.
• Opera Aria Opera · Browser AI · 2023 — AI assistant in Opera browser.
• Opus Clip Opus Clip · Video Editing · 2023 — AI clip extraction from long videos.
• Otter.ai Otter.ai · Transcription · 2023 — AI meeting transcription and notes.
• Pyramid Flow Peking University / Kuaishou · Video Generation · 2024 — Open-source text-to-video via pyramidal flow matching: generates low-res then progressively adds detail for efficiency. Runs on ~16GB VRAM. MIT license. Academic collaboration (PKU / Kuaishou / others).
• Perceptron Isaac 0.1 Perceptron · LLM · 2025 — Perceptron's first perceptive-language model: an open 2B-parameter model for real-world visual understanding, from the team behind Meta's Chameleon; Perceptron says it matches models over 50x its size. CC BY-NC 4.0.
• Perplexity Pro Perplexity AI · Search AI · 2024 — Perplexity Pro, Perplexity's paid plan for individuals at $17 a month billed annually: it raises the free plan's limits on Pro Search, file uploads and image generation, adds limited video generation and lets subscribers choose a preferred AI model per query, and includes Deep Research, Perplexity Computer access with 4,000 bonus credits, Model Council and premium data such as PitchBook and Statista. Perplexity Max, at $167 a month billed annually, raises the limits further.
• Perplexity R1 1776 Perplexity AI · LLM · 2025 — DeepSeek-R1 post-trained by Perplexity to remove Chinese Communist Party censorship while keeping its reasoning ability; released with open weights under the MIT license in February 2025. Perplexity has since removed its own Hugging Face repository, but community copies remain.
• Perplexity Sonar Perplexity AI · Search LLM · 2025 — Sonar, Perplexity's in-house answer model, rebuilt in February 2025 on Llama 3.3 70B and further trained for factual, readable answers grounded in search results: Perplexity said it beat GPT-4o mini and Claude 3.5 Haiku on user satisfaction and ran at 1,200 tokens per second on Cerebras hardware for Pro users. The Sonar API opened in January 2025; since Sonar Chat Completions support ended September 27, 2026, Sonar runs through the Agent API at $0.25/$2.50 per million tokens.
• Sonar Large Perplexity AI · Search LLM · 2025 — Sonar Large, the larger model in Perplexity's Llama-based Sonar API family, last offered as llama-3.1-sonar-large-128k-online, which answered questions using live web search with a 128K-token context. Perplexity retired the llama-3.1-sonar models on February 22, 2025, pointing developers to its new Sonar and Sonar Pro models.
• PhotoRoom PhotoRoom · Image Editing · 2022 — AI background removal and product photos.
• Picsart AI Picsart · Image Editing · 2023 — AI editing tools in Picsart suite.
• Pictory Pictory · Video Editing · 2022 — AI video editing. Long-form to short-form.
• Pika 1.0 Pika Labs · Video Generation · 2023 — Text/image to video. Effects and editing.
• Pika 2.0 Pika Labs · Video Generation · 2024 — Text/image-to-video. Lipsyncing. Scene extensions. Beginner-friendly.
• Pixlr AI Pixlr · Image Editing · 2023 — AI-powered photo editor.
• PixVerse PixVerse · Video Generation · 2024 — Text-to-video with character consistency.
• Play.ht Play.ht · Speech Synthesis · 2023 — AI voice generator with 900+ voices. Meta acquired PlayAI in July 2025, and the Play.ht service has since shut down.
• Playground v2.5 Playground AI · Image Generation · 2024 — Playground's v2.5 diffusion text-to-image model, generating 1024x1024 images plus portrait and landscape aspect ratios with the SDXL architecture (two fixed text encoders, OpenCLIP-ViT/G and CLIP-ViT/L); Playground reports its user studies preferred it over SDXL, Playground v2, PixArt-α, DALL-E 3 and Midjourney 5.2 for aesthetic quality.
• Pleias 1.0 PleIAs · LLM · 2025 — Family of small base models (350M to 3B) trained from scratch only on open data that is out of copyright or permissively licensed (Common Corpus). Apache 2.0.
• Podcastle Podcastle · Audio Editing · 2023 — AI podcast editing and recording. Podcastle now operates as Async; podcastle.ai redirects to async.com.
• Laguna S 2.1 Poolside · Code LLM · 2026 — Open-weight 118B-parameter MoE (8B active) built for agentic coding: up to 1M-token context, 70.2 pass@1 on Terminal-Bench 2.1 — matching or beating open models several times its size, including DeepSeek-V4-Flash and Nemotron 3 Ultra.
• Laguna XS 2.1 Poolside · Code LLM · 2026 — Smaller open sibling of Laguna S 2.1 (33B total / 3B active), released earlier in July 2026. Date recorded to first of month per AIDb convention.
• METAGENE-1 Prime Intellect · Science/Biology · 2025 — METAGENE-1 is an AI model or product entry requiring row-level review for exact capabilities and positioning.
• Prime Intellect INTELLECT-1 Prime Intellect · LLM · 2024 — First collaboratively trained 10B model, trained from scratch on 1T tokens across up to 14 concurrent nodes on three continents with compute from 30 contributors. Apache 2.0.
• Prime Intellect INTELLECT-2 Prime Intellect · LLM · 2025 — 32B model trained through a globally distributed, permissionless reinforcement-learning run on community-contributed GPUs (prime-rl framework, GRPO with verifiable rewards). Apache 2.0.
• Lensa AI Prisma Labs · Image Generation · 2022 — AI avatar generation from selfies. Viral.
• QuillBot QuillBot · AI Writing · 2017 — AI paraphrasing and grammar tool.
• Quizlet Q-Chat Quizlet · Education AI · 2023 — AI study tutor.
• Poe Quora · Chat Interface · 2023 — Multi-model AI chat platform by Quora.
• Recall AI Recall.ai · Meeting AI · 2023 — AI meeting bot API. Record and transcribe.
• Recraft V3 Recraft · Image Generation · 2024 — Design-focused image generation. Vector/brand.
• Recraft V4 Recraft · Image Generation · 2026 — Ground-up rebuild of Recraft's image model for design workflows: raster and native vector (SVG) output, 1-megapixel (V4 Pro at 4MP for print), plus vectorization, style reference and natural-language editing.
• Reflect AI Reflect · Productivity AI · 2023 — Networked note-taking app with daily notes and backlinked notes. Reflect now ships as Reflect Open, an open-source (MIT) app that keeps notes as local Markdown files synced through iCloud Drive or Git; its optional AI features use your own API key. Reflect Classic is still maintained.
• Reka Flash 3 Reka AI · LLM · 2025 — 21B general-purpose reasoning model trained from scratch; Reka says it is competitive with OpenAI o1-mini and well suited to low-latency or on-device use. Apache 2.0.
• Reka Flash 3.1 Reka AI · LLM · 2025 — Update to Reka Flash 3 that is especially strong at coding and as a base for agentic fine-tuning; Reka reports a 10-point gain on LiveCodeBench v5 from large-scale reinforcement learning. Apache 2.0.
• Relevance AI Relevance AI · AI Platform · 2023 — No-code AI agent builder.
• Renmin YuLan-Mini Renmin University · LLM · 2025 — Lightweight 2.4B base model from Renmin University that rivals models trained on far more data despite pre-training on only 1.08T tokens; strongest in math and code. MIT license.
• Replit Agent Replit · Code Agent · 2024 — Replit Agent, Replit's AI agent that builds apps and websites from a chat description, with no coding experience required: it searches the web when needed, offers extended thinking and high-power models for harder tasks, wires in Replit's built-in database and authentication plus services such as Stripe and OpenAI, and tests its work in a browser, fixing issues in a reflection loop. It can also build other agents and automations; Core costs $20 a month and Pro, with 10 parallel agents, $100.
• Replit Code Replit · Code LLM · 2023 — Replit Code v1.5, a 3.3B-parameter code-completion model trained on 1T tokens of permissively licensed code in 30 programming languages from BigCode's Stack Dedup plus developer-focused Markdown, reStructuredText and StackExchange text; 4,096-token context and a custom 32,768-token vocabulary.
• Resemble AI Resemble AI · Speech Synthesis · 2023 — Deepfake-detection and AI-security platform: detection for audio, images and video, plus watermarking, identity verification and audio source tracing. Formerly a voice-cloning and text-to-speech company; its Chatterbox voice models remain free and open, but voice is no longer a paid product.
• Reve 2.0 Reve · Image Generation · 2026 — Jumped to #2 on Text-to-Image Arena (~1200 Elo) with native 4K output, code-like layout control and precise editing; the layout-first editor is the differentiator for graphic-iteration workflows. Early tests found inconsistent portrait identity.
• Reve 2.1 Reve · Image Generation · 2026 — Layer-based generation: images are built through a layout engine so every element lands on its own editable layer — edit one element and the image rebuilds around it. #2 on Text-to-Image Arena at 1306 (28 points clear of the field), dethroning Meta's Muse Image after ~30 hours; #8 on single-image editing (par with Nano Banana Pro); improved prompt understanding, world knowledge and foreign-text rendering.
• Allegro Rhymes AI · Video Generation · 2024 — Open-source text-to-video model, ~6s clips at 720p/15fps. Apache 2.0. Runs on consumer GPUs (~24GB VRAM, less when quantized). From Rhymes AI (also maker of Aria multimodal LLM).
• Riffusion Riffusion · Music Generation · 2022 — Music generation from spectrogram diffusion. The Riffusion team rebranded as ProducerAI, which Google acquired in February 2026; riffusion.com now serves Google Flow Music.
• Riverside.fm Riverside · Recording · 2023 — AI-powered podcast/video recording.
• Ruliad Deepthought-8B Ruliad · LLM · 2025 — Small reasoning model built on Llama 3.1 8B that lays out its reasoning as structured JSON steps to make it more transparent and controllable.
• Runway (platform) Runway · AI Platform · 2023 — Runway, the company building what it calls 'Real-World Intelligence' models that understand, simulate and act in the world, offered through three platforms: Runway Creative, a cloud suite for generating and editing video, images and audio that Runway says 60 million+ creatives use; Runway Dev, an API platform for models from Runway and other labs; and Runway Robotics, simulation and policy tools built on its GWM-1 General World Model. It bills Gen-4.5 as the world's best video model.
• Runway Gen-1 Runway · Video Generation · 2023 — Runway's Gen-1 (February 2023), a video-to-video model that synthesizes new videos by applying the composition and style of an image or text prompt to the structure of a source video, with stylization, storyboard, mask, render and customization modes; in Runway's user studies its results were preferred over Stable Diffusion 1.5 (73.53%) and Text2Live (88.24%) for image-to-image and video-to-video translation. Runway deprecated it on April 6, 2025.
• Runway Gen-2 Runway · Video Generation · 2023 — Runway's Gen-2 (2023), a multimodal AI system that generates novel videos from text, images or video clips: it added Text to Video, Text + Image to Video and Image to Video modes to Gen-1's video-to-video modes for stylization, storyboards, masks, renders and customization. Runway began revoking access to Gen-2 in mid-March 2025 and it has been unavailable to all users since May 11, 2025, with Gen-4.5 named as the replacement.
• Runway Gen-3 Alpha Runway · Video Generation · 2024 — Runway's Gen-3 Alpha (June 17, 2024), the first of its next-generation models trained on new infrastructure for large-scale multimodal training, jointly on videos and images, with major gains in fidelity, consistency and motion over Gen-2; temporally dense captions enabled imaginative transitions and precise key-framing, and it powered Text to Video, Image to Video, Motion Brush, Advanced Camera Controls and Director Mode, shipping with C2PA provenance. Runway retired it on July 8, 2026.
• Runway Gen-3 Alpha Turbo Runway · Video Generation · 2024 — Runway's Gen-3 Alpha Turbo, released to all plans on August 14, 2024 as a version of Gen-3 Alpha that Runway said generated 7x faster for half the price while matching its performance in many use cases; it later gained extensions up to 34 seconds, video to video, first, middle and last keyframes, camera control and Act-One, and in September 2024 it became the first model offered through the Runway API. Runway retired it on July 30, 2026.
• Runway Gen-4 Runway · Video Generation · 2025 — Runway's Gen-4 video model, built for world consistency: from a single reference image plus instructions it keeps characters, objects and locations consistent across scenes, lighting and camera angles without fine-tuning, and Runway says it produces highly dynamic, realistic motion with strong prompt adherence and marks a step toward simulating physics. It reached paid plans in April 2025, followed by the faster Gen-4 Turbo and Gen-4 References for consistent characters and locations.
• Runway Gen-4.5 Runway · Video Generation · 2025 — Codename 'Whisper Thunder': #1 on AA Text-to-Video at launch (1247 Elo) with world consistency — characters, environments and objects coherent across cuts; Motion Brush 2.0 + Camera Director; co-optimized with NVIDIA Hopper/Blackwell. Rolled out late Nov-Dec 2025; since displaced at #1 but retains the strongest control surface.
• Stable Diffusion (LDM-KL-8-G) Runway · Image Generation · 2022 — Stable Diffusion, the latent diffusion text-to-image model released publicly on August 22, 2022. Its V1 release was led by Robin Rombach (Stability AI) and Patrick Esser (Runway) of LMU Munich's CompVis group, building on their Latent Diffusion Models paper, which runs diffusion in a pretrained autoencoder's latent space with text cross-attention to cut compute. The v1.4 weights shipped under CreativeML OpenRAIL-M, allowing commercial use, and needed about 6.9 GB of VRAM.
• RWKV-6 RWKV Foundation · LLM · 2024 — RWKV-6 (Finch), an RNN-style sequence model improving on RWKV-4 with multi-headed matrix-valued states and a dynamic recurrence mechanism that add expressivity while keeping RNN-like inference efficiency; Finch models with 1.6 and 3.1 billion parameters were trained alongside a new 1.12-trillion-token multilingual corpus and released on Hugging Face under Apache 2.0.
• RWKV-7 Goose RWKV Foundation · LLM · 2025 — RNN-style sequence architecture with constant memory use and constant per-token inference time; its 2.9B model set a new 3B-class state of the art on multilingual tasks despite far fewer training tokens.
• LSTM LM RWTH Aachen · LLM · 1997 — LSTM language model: Sundermeyer, Schlüter and Ney at RWTH Aachen (Interspeech 2012) applied the long short-term memory network (Hochreiter and Schmidhuber, 1997) to language modeling, improving perplexity by about 8% over standard recurrent-network LMs and lowering speech-recognition error rates.
• Rytr Rytr · AI Writing · 2021 — AI content writing assistant.
• Sail Sailor2 SAIL (Sea AI Lab) · LLM · 2025 — Community-driven multilingual LLMs for Southeast Asia in 1B, 8B and 20B sizes, built on Qwen 2.5. Apache 2.0.
• Sailor2 SAIL (Sea AI Lab) · LLM · 2024 — Southeast Asian language model.
• BLIP-2 Salesforce · Multimodal · 2023 — Salesforce's BLIP-2 vision-language model, which links a frozen CLIP-like image encoder and a frozen large language model (here OPT-2.7b) through a trained Querying Transformer (Q-Former); used for image captioning, visual question answering and chat about images.
• CodeGen 2.5 Salesforce · Code LLM · 2023 — Salesforce's CodeGen2.5-7B program-synthesis models, trained on StarCoderData for 1.4T tokens, which Salesforce reports are competitive with StarCoderBase-15.5B at under half the size; they support infilling and many programming languages, in multi and Python-tuned mono versions (Apache 2.0) and a research-only instruct version.
• ProGen Salesforce · Science/Biology · 2023 — Salesforce's ProGen, a 1.2B-parameter language model trained on about 280 million protein sequences conditioned on taxonomic and keyword tags (such as molecular function and cellular component), letting it generate protein sequences with fine-grained control as measured by sequence similarity, secondary-structure accuracy and conformational energy.
• Salesforce Einstein GPT Salesforce · Enterprise AI · 2023 — Generative AI for CRM.
• SambaNova EvaByte SambaNova Systems · LLM · 2025 — Tokenizer-free byte-level LLM (6.5B) developed with University of Hong Kong. Trained on 1.5T bytes; multibyte prediction for faster decoding. Open weights.
• Samsung Galaxy AI Samsung · AI Platform · 2024 — On-device AI features for Galaxy phones.
• GigaChat Sber · Chat Interface · 2023 — Sber's Russian language AI chatbot.
• Kandinsky 3 Sber · Image Generation · 2023 — Sber AI and AIRI's Kandinsky 3.0, an open text-to-image latent diffusion model with 11.9 billion parameters that, compared with Kandinsky 2.x, uses a U-Net twice as large and a text encoder ten times larger and drops the diffusion mapping, improving text understanding and specific domains in side-by-side human comparisons; it adds inpainting, super-resolution, image editing and image-to-video, and the distilled Kandinsky 3.1 generates in 4 steps, 20 times faster.
• Scribe AI Scribe · Documentation AI · 2023 — Auto-generate step-by-step guides.
• ServiceNow Apriel-5B ServiceNow · LLM · 2025 — First model in ServiceNow's Apriel family: a 5B decoder-only transformer trained on 4.5T+ tokens, built for high throughput and efficiency. MIT license.
• Sesame CSM Sesame AI · Voice / TTS · 2025 — Conversational Speech Model. Human-like voice with emotions, pauses, laughter.
• Agents-A1 Shanghai AI Lab · LLM · 2026 — Apache 2.0 agentic 35B MoE built on Qwen3.5-35B-A3B by the InternScience team, trained specifically for long-horizon agent work with a 256K context window.
• InternLM 2.5 Shanghai AI Lab · LLM · 2024 — InternLM2.5, open 7B base and chat models with math reasoning that their developers say surpasses Llama 3 and Gemma 2 9B, a 1M-token context window with near-perfect needle-in-a-haystack retrieval, and stronger tool use, including gathering information from more than 100 web pages through MindSearch.
• Intern-S1 Shanghai AI Laboratory/SenseTime · Reasoning LLM · 2025 — Intern-S1, the InternLM team's open (Apache 2.0) scientific multimodal reasoning model: a mixture-of-experts model with 241B total and 28B active parameters, built on Qwen3's 235B MoE language model and a 6B InternViT vision encoder, continually pre-trained on 5 trillion tokens, over 2.5 trillion of them scientific, with a dynamic tokenizer for molecular formulas, protein sequences and seismic signals. Its card reports 83.5 on MMLU-Pro, 77.7 on MMMU, 86.0 on AIME 2025 and 83.4 on ChemBench.
• InternLM3 Shanghai AI Laboratory/SenseTime · LLM · 2025 — InternLM3-8B-Instruct (January 2025), the InternLM team's open (Apache 2.0) 8B instruction model for general use and advanced reasoning, with a deep-thinking mode using long chain-of-thought and a normal response mode; trained on only 4 trillion high-quality tokens, which the team says cut training cost by more than 75% versus similar-size models, it scores 83.1 on CMMLU, and the team says it surpasses Llama 3.1 8B and Qwen2.5 7B on reasoning and knowledge-intensive tasks.
• InternVL 2.5 Shanghai AI Laboratory/SenseTime · LLM · 2024 — InternVL 2.5 (December 2024), OpenGVLab's open multimodal LLM series from 1B to 78B parameters that keeps InternVL 2.0's 'ViT-MLP-LLM' design while improving training, test-time strategies and data quality; the 78B flagship pairs the InternViT-6B vision encoder with Qwen2.5-72B-Instruct. Its paper reports performance rivaling GPT-4o and Claude 3.5 Sonnet and calls it the first open-source multimodal model to pass 70% on MMMU, with chain-of-thought reasoning adding 3.7 points.
• OREAL-32B Shanghai AI Laboratory/SenseTime · LLM · 2025 — OREAL-32B, an open (Apache 2.0) 32B math-reasoning model from the InternLM team, trained with OREAL (Outcome REwArd-based reinforcement Learning), an RL method for tasks with only right-or-wrong rewards that clones best-of-N correct samples, reshapes rewards for wrong ones and uses a token-level reward model to find key reasoning steps; it scores 95.0 pass@1 on MATH-500, above earlier distillation-trained 32B models, and 60.0 on AIME 2024. A 7B sibling and SFT versions are also open.
• Vidu (Shengshu) Shengshu AI · Video Generation · 2024 — Chinese video generation model.
• Shopify Sidekick Shopify · Commerce AI · 2023 — AI assistant for Shopify merchants.
• Amper Music Shutterstock · Music Generation · 2020 — AI music creation platform. Acquired by Shutterstock.
• Luminar Neo AI Skylum · Image Editing · 2022 — AI photo editor. Sky replacement, relight.
• SkyReels V1 Skywork AI · Video Generation · 2025 — Open-source video model fine-tuned from HunyuanVideo on 10M+ film/TV clips; emphasis on realistic humans and cinematic facial expressions. V2 (Diffusion Forcing) lowered hardware needs and extended length.
• Slack AI Slack/Salesforce · Productivity AI · 2024 — AI summaries and search in Slack.
• Snapchat My AI Snap · Chat Interface · 2023 — AI chatbot in Snapchat, powered by large language models including OpenAI's GPT and Google's Gemini models.
• Snowflake Arctic Snowflake · LLM · 2024 — 480B MoE. Enterprise focused.
• Snowflake Arctic Embed Snowflake · Embeddings · 2024 — Snowflake's Arctic-embed text embedding models for retrieval, in five sizes from xs (22M parameters, 384 dimensions) to l (335M parameters, 1,024 dimensions); Snowflake now recommends its v2.0 models (December 2024), built for multilingual workloads.
• Soundraw Soundraw · Music Generation · 2023 — AI music for creators. Customizable.
• Sourcegraph Cody Sourcegraph · Code AI · 2023 — AI coding assistant with codebase context. Cody Free and Pro ended on July 23, 2025; Cody Enterprise continues.
• Speechify Speechify · Speech Synthesis · 2023 — Text-to-speech reader. 200+ voices.
• Splash Music Splash · Music Generation · 2023 — Music creation tools and games.
• Squarespace Blueprint AI Squarespace · Web Design · 2024 — AI-guided website creation.
• Clipdrop Stability AI · Image Editing · 2023 — Clipdrop, a web app of AI image tools, including text to image, background removal and replacement, Cleanup for erasing objects, people and text, uncrop, 2x or 4x upscaling, relighting and resizing, with an API for developers. It was built by InitML, founded in 2020 by Cyril Diagne, Jonathan Blanchet and Damien Henry; Stability AI acquired InitML in February 2023 and sold it to Jasper in February 2024, and Clipdrop's site now points developers to the Jasper API.
• DreamStudio Stability AI · Image Generation · 2022 — Stability AI's former web app for Stable Diffusion image generation; dreamstudio.ai now redirects to Stability AI's Brand Studio.
• Japanese StableLM Stability AI · LLM · 2023 — Stability AI's Japanese StableLM; its first model, Japanese-StableLM-Base-Alpha-7B, is a 7B-parameter decoder-only model on the GPT-NeoX architecture, pretrained on about 750B tokens of Japanese and English text (Wikipedia, Japanese mC4, CC-100 and OSCAR, and RedPajama) with a 2,048-token context, alongside an instruction-tuned Instruct-Alpha-7B.
• SDXL Turbo Stability AI · Image Generation · 2023 — Stability AI's SDXL Turbo, a version of SDXL 1.0 distilled with Adversarial Diffusion Distillation to generate 512x512 images from a text prompt in 1 to 4 sampling steps, fast enough for real-time synthesis; Stability reports human raters preferred its single-step images over LCM-XL's four-step output for image quality and prompt following.
• Stability Audio Open Stability AI · Music Generation · 2024 — Stability AI's Stable Audio Open 1.0, an open-weights text-to-audio model with about 1.2B parameters, trained on Creative Commons-licensed audio, that generates 44.1kHz stereo sound; Stability reports performance competitive with the state of the art across metrics, with strong realism scores (FDopenl3).
• Stable Audio Stability AI · Music Generation · 2023 — Stability AI's Stable Audio (September 13, 2023), its first product for music and sound generation: a web app that turns text prompts plus a chosen length into 44.1 kHz music and sound effects using latent diffusion conditioned on text metadata, audio duration and start time, trained on AudioSparx music and metadata. At launch the free tier made tracks of up to 45 seconds and Pro made 90-second tracks for commercial use; Stable Audio 2.0 (April 2024) extended it to three-minute tracks.
• Stable Audio 2.0 Stability AI · Music Generation · 2024 — Stability AI's Stable Audio 2.0 (April 3, 2024), a latent diffusion model that generates full tracks of up to three minutes in 44.1 kHz stereo with coherent structure (intro, development and outro) from a single text prompt, and adds audio-to-audio generation that transforms uploaded samples, plus sound effects and style transfer; it pairs a highly compressed autoencoder with a diffusion transformer like Stable Diffusion 3's and was trained only on licensed AudioSparx music, honoring opt-outs.
• Stable Cascade Stability AI · Image Generation · 2024 — Stability AI's Stable Cascade text-to-image model, built on the Würstchen architecture as a three-stage cascade (Stages A, B and C) that works in a 42x-compressed latent space (a 1024x1024 image encodes to 24x24), for faster inference and cheaper training than Stable Diffusion; supports fine-tuning, LoRA, ControlNet, IP-Adapter and LCM.
• Stable Diffusion 1.5 Stability AI · Image Generation · 2022 — Most popular open image model. 860M params.
• Stable Diffusion 2.0 Stability AI · Image Generation · 2022 — Stability AI's Stable Diffusion 2.0 (November 24, 2022), an open-source release of text-to-image models using LAION's new OpenCLIP text encoder, generating 512x512 and 768x768 images, trained on an aesthetic subset of LAION-5B filtered for adult content; it added a 4x Upscaler Diffusion model enabling 2048x2048 output, depth2img for depth-guided image-to-image, and a new text-guided inpainting model, all designed to run on a single GPU.
• Stable Diffusion 2.1 Stability AI · Image Generation · 2022 — Stability AI's Stable Diffusion 2.1 (December 7, 2022), a fine-tune of Stable Diffusion 2.0 released a few weeks after it, with the same OpenCLIP text encoder; trained with more data and a less aggressive adult-content filter that cut false positives, it renders people better than 2.0, improves anatomy, hands and art styles, handles non-standard resolutions and extreme aspect ratios, and works better with negative prompts.
• Stable Diffusion 3 Stability AI · Image Generation · 2024 — Stability AI's Stable Diffusion 3, text-to-image models built on the Multimodal Diffusion Transformer (MMDiT), with separate weights for image and text tokens that share information both ways, trained by rectified flow for better text comprehension and typography. Its first open release, the 2B-parameter SD3 Medium (June 12, 2024), handled long prompts, hands, faces and in-image text on consumer GPUs; Stability later said it fell short of its standards and followed with SD 3.5 in October 2024.
• Stable Diffusion 3.5 Stability AI · Image Generation · 2024 — Stability AI's Stable Diffusion 3.5 (October 22, 2024), an open release of text-to-image models: SD 3.5 Large (8.1B parameters, 1 megapixel), Large Turbo (a distilled version that generates in 4 steps) and, from October 29, Medium (2.5B, an improved MMDiT-X architecture, 0.25 to 2 megapixels, about 9.9 GB of VRAM). Query-Key Normalization eases fine-tuning; the models run on consumer hardware, and the Stability AI Community License makes them free for commercial use below $1M in annual revenue.
• Stable Diffusion XL Stability AI · Image Generation · 2023 — Stability AI's SDXL 1.0 (July 26, 2023), an open text-to-image model with a 3.5B-parameter base model and a 6.6B-parameter ensemble pipeline in which a refiner model handles the final denoising steps; it generates native 1024x1024 images with better color, contrast and lighting than its predecessor, renders hands, text and spatial compositions better, needs only short prompts, runs on consumer GPUs with 8 GB of VRAM and supports LoRA fine-tuning. Released under CreativeML OpenRAIL++-M.
• Stable Diffusion XL (SDXL) Stability AI · Image Generation · 2023 — Stability AI's SDXL 1.0 latent diffusion text-to-image model, which uses two fixed text encoders (OpenCLIP-ViT/G and CLIP-ViT/L); the base model runs on its own or hands its latents to a separate refiner model for the final denoising steps.
• Stable Video Diffusion Stability AI · Video Generation · 2023 — Stability AI's Stable Video Diffusion (November 21, 2023), its first foundation model for generative video, based on Stable Diffusion: two image-to-video models generating 14 and 25 frames at customizable frame rates from 3 to 30 frames per second, which Stability said beat leading closed models in external user-preference studies. It launched as a research preview with code on GitHub and weights on Hugging Face, adaptable to tasks such as multi-view synthesis from a single image.
• Stable Zero123 Stability AI · 3D Generation · 2023 — Stability AI's Stable Zero123 view-conditioned image generation model, based on Zero123, which with Score Distillation Sampling turns a single input image into a 3D model; Stability reports better results than Zero123 and Zero123-XL.
• StableCode Stability AI · Code LLM · 2023 — Stability AI's StableCode-Completion-Alpha-3B: a 2.8B-parameter decoder-only model for single- and multi-line code completion with a 16K-token context, pretrained on the most-used languages in the 2023 Stack Overflow developer survey (300B tokens at a 4K context, then 200B at 16K).
• StableLM Stability AI · LLM · 2023 — Stability AI's StableLM language-model series, which began on April 20, 2023 with StableLM-Alpha 3B and 7B base models (CC BY-SA 4.0) and tuned chat versions; Stability later released improved Alpha v2 models (August 2023, trained on 1.1T tokens) and StableLM-3B-4E1T (September 2023, 1T tokens for 4 epochs), which it reported as state of the art among open 3B models at the time.
• StableLM 2 Stability AI · LLM · 2024 — Stability AI's Stable LM 2 language models; the 1.6B model is a decoder-only transformer (1.64B parameters, 24 layers, 4,096-token context) pretrained for two epochs on 2 trillion tokens of multilingual and code data.
• SV3D Stability AI · 3D Generation · 2024 — Stability AI's Stable Video 3D (SV3D, March 18, 2024), built on Stable Video Diffusion with camera-path conditioning, generates consistent novel views of an object from a single image: SV3D_u makes orbital videos without camera conditioning, and SV3D_p follows specified camera paths. Stability reports better quality and multi-view consistency than Stable Zero123 and Zero123-XL, and SV3D can produce 3D meshes through NeRF and mesh optimization with a masked score distillation loss.
• TripoSR Stability AI/Tripo · 3D Generation · 2024 — TripoSR, an open-source model from Tripo AI and Stability AI for fast feedforward 3D reconstruction from a single image, based on the Large Reconstruction Model (LRM) approach; it produces a 3D model in under 0.5 seconds on an NVIDIA A100, and its developers report it beats other open-source alternatives on public datasets. Code and weights are MIT-licensed.
• bolt.new StackBlitz · Code AI · 2024 — bolt.new, StackBlitz's AI builder for websites, apps and prototypes made by chatting, starting from a prompt, a Figma file, a GitHub repo or a template: Bolt routes each task to a suitable model, tests, refactors and iterates to cut errors (it claims 98% fewer) and handles projects 1,000 times larger than before, while Bolt Cloud adds hosting with custom domains and analytics, unlimited databases, authentication and SEO tools. Plans: Free, Pro at $25 a month and Teams at $30 per member.
• Alpaca Stanford · LLM · 2023 — Stanford's Alpaca 7B, Meta's LLaMA 7B fine-tuned on 52K instruction-following examples generated with a modified Self-Instruct method; in a preliminary human evaluation it behaved similarly to OpenAI's text-davinci-003. Released for research use only as a weight diff with its data and code; the live demo is suspended.
• ColBERT v2 Stanford · Retrieval · 2022 — Stanford's ColBERT (v2), a fast, accurate retrieval model for scalable BERT-based search over large text collections in tens of milliseconds: it encodes passages as matrices of token-level embeddings and matches queries through contextual late interaction (MaxSim), beating single-vector models while scaling to large corpora; ColBERTv2 (NAACL 2022) made late interaction lightweight.
• GloVe Stanford · Word Vectors · 2014 — Global Vectors for word representation.
• Stanford s1-32B Stanford · Reasoning LLM · 2025 — Reasoning model fine-tuned from Qwen2.5-32B-Instruct on just 1,000 examples; it matches o1-preview and scales at test time via 'budget forcing'. Apache 2.0.
• StarryAI StarryAI · Image Generation · 2021 — AI art generator app.
• Step-Video-T2V StepFun · Video Generation · 2025 — 30B-parameter text-to-video foundation model, up to 204 frames. Deep-compression Video-VAE (16x16 spatial, 8x temporal). Bilingual (EN/ZH) encoders, 3D full-attention DiT, Video-DPO. MIT license. Turbo + TI2V (image-to-video) variants.
• StepFun Step 3 StepFun · Reasoning LLM · 2025 — Multimodal reasoning mixture-of-experts with 321B total and 38B active parameters, co-designed (MFA attention, attention-FFN disaggregation) to cut decoding cost. Apache 2.0.
• Stripe Copilot Stripe · Finance AI · 2024 — AI for payment operations.
• Sudowrite Sudowrite · AI Writing · 2021 — AI fiction writing assistant.
• Bark Suno AI · Speech Synthesis · 2023 — Suno's Bark, a transformer-based text-to-audio model that generates realistic multilingual speech plus music, background noise, simple sound effects and nonverbal sounds like laughing, sighing and crying; it detects the language from the text, offers 100+ speaker presets and has been MIT-licensed for commercial use since May 2023. Suno built it for research and warns it can deviate unexpectedly from prompts.
• Suno v3 Suno AI · Music Generation · 2024 — Text-to-song. Vocals and instrumentals. 2-min.
• Suno v4 Suno AI · Music Generation · 2024 — Full song generation with vocals and lyrics. Multiple genres. AI music leader.
• Suno v5.5 Suno AI · Music Generation · 2026 — Suno's AI music update adding custom voice models - users can generate songs in their own voice - plus improved vocal realism and song structure over v4.
• Superhuman AI Superhuman · Email AI · 2023 — AI email writing and summarization.
• Synthesia Synthesia · Video Generation · 2020 — AI avatar video creation. 140+ languages.
• Synthesia EXPRESSIVE-1 Synthesia · Video Generation · 2024 — Expressive avatar model. Natural gestures.
• Tabnine Tabnine · Code Completion · 2019 — AI code completion. IDE plugins. Now part of Tricentis; tabnine.com redirects to tricentis.com.
• Taskade AI Taskade · Productivity AI · 2023 — AI productivity and project management.
• Tavily Tavily · Search AI · 2024 — Search API optimized for AI agents.
• Falcon Technology Innovation Institute · LLM · 2023 — UAE open-source LLM. Government-backed. Competitive benchmarks at release.
• Hunyuan Chat Tencent · Chat Interface · 2023 — Tencent's consumer AI chatbot.
• Hunyuan T1 Tencent · LLM · 2025 — Tencent's reasoning model, built on its TurboS hybrid Mamba-Transformer mixture-of-experts base and released in March 2025. Tencent reported results on par with or slightly above DeepSeek-R1 on MMLU-Pro, C-Eval and AIME, priced at 1 yuan per million input tokens and 4 yuan per million output tokens.
• Hunyuan Turbo S Tencent · LLM · 2025 — Tencent's fast-response model, built to answer near-instantly rather than reason at length; Tencent said it is competitive with DeepSeek-V3 on common benchmarks. The TurboS line uses a hybrid Transformer-Mamba mixture-of-experts design (56B active of 560B total parameters in Tencent's technical report) and is offered through Tencent Cloud.
• Hunyuan Video Tencent · Video Generation · 2024 — Tencent's open-source HunyuanVideo text-to-video foundation model, with over 13 billion parameters, which Tencent called the largest open-source video model at its December 2024 release; Tencent reports its professional human evaluation rated it above Runway Gen-3, Luma 1.6 and three top Chinese video models.
• Hunyuan-A13B Tencent · LLM · 2025 — Hunyuan-A13B, Tencent's open-weight mixture-of-experts language model with 80B total and 13B active parameters: it offers both fast and slow thinking modes, a native 256K-token context, grouped query attention and FP8 and GPTQ-Int4 quantized versions, and is tuned for agent tasks, where Tencent reports leading results on BFCL-v3, τ-Bench and C3-Bench. Tencent released Pretrain and Instruct weights on Hugging Face with a technical report under the Tencent Hunyuan Community License.
• Hunyuan-DiT Tencent · Image Generation · 2024 — Tencent's Hunyuan-DiT, a multi-resolution diffusion transformer for text-to-image generation with fine-grained understanding of both Chinese and English prompts; Tencent reports it set a new state of the art in Chinese-to-image generation among open-source models in its human evaluation.
• Hy-MT2-1.8B Tencent · Translation LLM · 2026 — Compact 1.8B translation model, 33 language pairs, 8K context - the small sibling of Hy-MT2-30B-A3B.
• Hy-MT2-30B-A3B Tencent · Translation LLM · 2026 — Tencent Hunyuan translation MoE: 30B total / 3B active, 33 language pairs, 8K context. Successor to the Hy-MT translation model.
• YOLO-World Tencent · Vision/CV · 2024 — YOLO-World, an open-vocabulary object detector from Tencent AI Lab, Tencent's ARC Lab and Huazhong University of Science and Technology (CVPR 2024): it adds vision-language modeling to fast YOLO detectors so they find objects named in text prompts zero-shot, using a re-parameterizable vision-language path aggregation network and region-text contrastive pre-training, reaching 35.4 AP on LVIS at 52 FPS on a V100. Later updates added v2, high-resolution and image-prompt models; code is GPL-3.0.
• Tesla Optimus Gen 2 Tesla · Robotics · 2023 — Humanoid robot. Improved dexterity.
• TextCortex TextCortex · AI Writing · 2023 — AI writing assistant with personas.
• Arc Search The Browser Company · Browser AI · 2024 — AI-first browser. Browse for Me feature.
• Inkling Thinking Machines Lab · Multimodal LLM · 2026 — Thinking Machines' first-ever model: 975B total / 41B active MoE pretrained from scratch on 45T multimodal tokens (text, images, audio, video), released under Apache 2.0 — the most permissive license at that parameter scale. Encoder-free native reasoning over text, image and audio; 1M-token context; 41 on the AA Index (top US open-weights model at release).
• Inkling-Small Thinking Machines Lab · Multimodal LLM · 2026 — 276B / 12B-active open MoE, natively multimodal with an encoder-free architecture (hierarchical patch encoding for images, dMel spectrograms straight into the decoder for audio) and variable thinking effort. On-policy distillation from Inkling plus two extra weeks of agentic-coding RL let it beat its 975B teacher on SWE-Bench Verified (80.2 vs 77.6) and ARC-AGI-2 (40.1 vs 36.5), though factual recall regressed hard (SimpleQA 20.6 vs 43.9).
• Casetext CoCounsel Thomson Reuters · Legal AI · 2023 — AI legal research assistant.
• Falcon 2 TII · LLM · 2024 — TII's Falcon 2 release: Falcon2-11B, an 11B-parameter causal decoder-only base model trained on over 5,000B tokens of RefinedWeb plus curated corpora, in English and 10 European languages, with an 8,192-token context; it is under the TII Falcon License 2.0 (based on Apache 2.0), and vision comes from the separate Falcon2-11B-VLM.
• Falcon 3 TII · LLM · 2024 — UAE open-source LLM. Government-backed. Competitive benchmarks at release.
• Falcon-180B TII · LLM · 2023 — TII's Falcon-180B, the largest of its open Falcon causal decoder models (alongside 7B and 40B), trained on over 3.5 trillion tokens of mainly web data, the largest openly documented pretraining run at the time, on up to 4,096 A100 GPUs; TII reports it significantly outperforms PaLM and Chinchilla, improves on Llama 2 and Inflection-1 and nears PaLM-2-Large at lower cost, and it was released with a 600B-token extract of its web dataset.
• Falcon-Edge TII · LLM · 2025 — UAE open-source LLM. Government-backed. Competitive benchmarks at release.
• Falcon-H1 TII · LLM · 2025 — UAE open-source LLM. Government-backed. Competitive benchmarks at release.
• Falcon2-11B-VLM TII · Multimodal LLM · 2024 — TII's Falcon2-11B vision-language model: the Falcon2-11B chat model (trained on more than 5,000B tokens of RefinedWeb and curated corpora) paired with a CLIP ViT-L/14 vision encoder and high-resolution dynamic encoding for fine image detail; English.
• Tilde AI TildeOpen-30b Tilde AI · LLM · 2025 — 30B dense open LLM built for European languages, including smaller ones such as Latvian, Lithuanian and Icelandic; funded by the European Commission through EuroHPC. CC BY 4.0.
• tl;dv tl;dv · Transcription · 2023 — AI meeting recorder. Timestamp highlights.
• TNG DeepSeek Chimera TNG Technology Consulting · LLM · 2025 — TNG Technology Consulting's DeepSeek-R1T-Chimera, an open-weights model made by merging the weights of DeepSeek-R1 and DeepSeek-V3 (0324) into a DeepSeek-MoE transformer, aiming to pair R1's intelligence with V3's token efficiency; TNG has since released R1T2-Chimera, which it says is both faster and smarter than R1.
• TNG DeepSeek-TNG-R1T2-Chimera TNG Technology Consulting · LLM · 2025 — 671B 'Tri-Mind' model assembled from DeepSeek R1-0528, R1 and V3-0324 with TNG's Assembly-of-Experts method, as the successor to the original R1T Chimera. MIT license.
• Deepcoder-14B-Preview Together AI · Code LLM · 2025 — Code reasoning model released by the Agentica team and Together AI, fine-tuned from DeepSeek-R1-Distill-Qwen-14B with distributed reinforcement learning; scores 60.6% Pass@1 on LiveCodeBench v5, matching OpenAI's o3-mini (low) with 14B parameters. MIT-licensed open weights.
• RedPajama Together AI · LLM · 2023 — Open reproduction of LLaMA training data.
• RedPajama-INCITE-7B Together AI · LLM · 2023 — Together's RedPajama-INCITE-7B-Base, a 6.9B-parameter English base language model trained on the RedPajama-Data-1T dataset using 3,072 V100 GPUs from the INCITE 2023 program, built with partners including Ontocord.ai, ETH DS3Lab, MILA, Stanford CRFM, Hazy Research and LAION; instruction-tuned and chat versions are also available.
• Swallow Tokyo Inst. of Tech. · LLM · 2024 — Japanese-continued LLM training.
• Tome AI Tome · Presentation AI · 2023 — AI presentation generator, since discontinued: Tome's founders shut the app down to build Lightfield, an AI-native CRM.
• Topaz Photo AI Topaz Labs · Image Enhancement · 2022 — AI noise reduction, sharpening, upscaling.
• YOLOv10 Tsinghua · Vision/CV · 2024 — Tsinghua's YOLOv10, a real-time end-to-end object detector that drops non-maximum suppression post-processing through consistent dual assignments in training and redesigns YOLO components for efficiency; six sizes from N (2.3M parameters) to X (29.5M), and its authors report YOLOv10-S is 1.8x faster than RT-DETR-R18 at similar COCO accuracy.
• Depth Anything Tsinghua/HK University · Vision/CV · 2024 — Depth Anything (CVPR 2024), a monocular depth estimation model trained on 1.5M labeled images and more than 62M unlabeled images, released in Small (24.8M parameters), Base (97.5M) and Large (335.3M) versions; it also underpins a depth-conditioned ControlNet and became the default depth processor for InstantID and InvokeAI.
• Depth Anything V2 Tsinghua/HK University · Vision/CV · 2024 — Depth Anything V2, a monocular depth estimation model that its authors say is much better than V1 in fine detail and robustness, and faster, smaller and more accurate than Stable Diffusion-based depth models; released in Small (24.8M parameters), Base (97.5M) and Large (335.3M) versions plus metric-depth models, with a 1.3B Giant listed as coming soon.
• ProtTrans TU Munich · Science/Biology · 2021 — ProtTrans, protein language models: two autoregressive models (Transformer-XL, XLNet) and four autoencoder models (BERT, ALBERT, ELECTRA, T5) trained on up to 393 billion amino acids from UniRef and BFD using the Summit supercomputer and TPU pods; the best embeddings (ProtT5) were the first to beat the state of the art in protein secondary-structure prediction without evolutionary information.
• Koala UC Berkeley · LLM · 2023 — LLaMA fine-tuned on dialogue data.
• Starling-7B UC Berkeley · LLM · 2023 — RLAIF-tuned model. Strong chat.
• Starling-LM 7B UC Berkeley · LLM · 2023 — Starling Language Model 7B variant.
• Udio Udio · Music Generation · 2024 — AI music generation. High quality production. Competitor to Suno.
• Udio v2 Udio · Music Generation · 2025 — Improved audio quality and style control.
• YOLOv5 Ultralytics · Vision/CV · 2020 — Ultralytics' YOLOv5, a fast, accurate and easy-to-use PyTorch computer vision model for object detection, instance segmentation and image classification; Ultralytics calls it mature and production-proven and points users who need newer architectures or tasks such as pose estimation to its ultralytics package.
• YOLOv8 Ultralytics · Vision/CV · 2023 — Ultralytics' real-time vision model, released January 10, 2023, with pretrained versions for object detection, instance segmentation, classification, pose estimation and oriented object detection. No longer Ultralytics' latest YOLO: the company now points new projects to YOLO26.
• Unitree G1 Unitree · Robotics · 2024 — Affordable humanoid robot platform.
• Feedforward NN University of Montreal · LLM · 1986 — Feedforward neural language model: Bengio, Ducharme, Vincent and Jauvin at the Université de Montréal (A Neural Probabilistic Language Model, JMLR 2003) learned a distributed vector for each word jointly with a network that predicts the next word, the forerunner of modern word embeddings.
• VGG16 University of Oxford · Vision/CV · 2014 — The University of Oxford Visual Geometry Group's VGG16, the 16-weight-layer network from Karen Simonyan and Andrew Zisserman's very deep convolutional networks study, which showed that stacking very small 3x3 convolution filters to a depth of 16 to 19 layers markedly beat earlier designs; it underpinned the team's ILSVRC-2014 entry, which took first place in localisation and second in classification, and its weights were released under a Creative Commons Attribution license.
• VGG19 University of Oxford · Vision/CV · 2014 — The University of Oxford Visual Geometry Group's VGG19, the 19-weight-layer network from Karen Simonyan and Andrew Zisserman's very deep convolutional networks study, which showed that stacking very small 3x3 convolution filters to a depth of 16 to 19 layers markedly beat earlier designs; it underpinned the team's ILSVRC-2014 entry, which took first place in localisation and second in classification, and its weights were released under a Creative Commons Attribution license.
• AlexNet University of Toronto · Vision/CV · 2012 — AlexNet. University of Toronto Vision/CV model.
• YOLOv3 University of Washington · Vision/CV · 2018 — The University of Washington's YOLOv3, an incremental update to the YOLO real-time object detector with a slightly bigger, more accurate network: at 320x320 it runs in 22 ms at 28.2 mAP, as accurate as SSD but three times faster, and reaches 57.9 mAP@50 in 51 ms on a Titan X versus 57.5 in 198 ms for RetinaNet. All code is online.
• LLaVA 1.5 University of Wisconsin · Multimodal LLM · 2023 — LLaVA-1.5 from the University of Wisconsin-Madison and Microsoft Research, an improved version of the open LLaVA (Large Language and Vision Assistant) multimodal model that connects a CLIP ViT-L/336px vision encoder to the Vicuna language model through an MLP projection and adds academic VQA training data; the 13B model set state-of-the-art results on 11 benchmarks using only 1.2M public training samples and trains in about a day on one 8-A100 node, and 7B and 13B versions were released.
• LLaVA 1.6 University of Wisconsin · Multimodal LLM · 2024 — LLaVA-1.6, released as LLaVA-NeXT on January 30, 2024, the successor to LLaVA-1.5 with 4x more image pixels (up to 672x672, 336x1344 or 1344x336) and better visual reasoning, OCR and world knowledge from an improved instruction-tuning mix of under 1M samples; its authors report the 34B version beats Gemini Pro on several benchmarks and trains in about a day on 32 A100s, alongside Vicuna-7B, Vicuna-13B and Mistral-7B versions.
• LLaVA-NeXT University of Wisconsin · Multimodal LLM · 2024 — LLaVA-NeXT (LLaVA-1.6), released January 30, 2024 by the LLaVA team as the successor to LLaVA-1.5, with 4x more image pixels (up to 672x672, 336x1344 or 1344x336) and better visual reasoning, OCR and world knowledge from an improved instruction-tuning mix of under 1M samples; its authors report the 34B version beats Gemini Pro on several benchmarks and trains in about a day on 32 A100s, alongside Vicuna-7B, Vicuna-13B and Mistral-7B versions.
• SOLAR 10.7B Upstage · LLM · 2023 — Upstage's depth-upscaled model.
• Upstage Solar Pro Upstage · LLM · 2024 — Upstage's commercial Solar Pro LLM line, sold through the Upstage API. The current version, Solar Pro 4, is an agentic LLM built for tool calling, terminal tasks and long-document reasoning; Solar Pro 3 and Solar Pro 2 end service on Oct 30, 2026.
• RFdiffusion UW/Baker Lab · Science/Biology · 2023 — RFdiffusion, an open-source diffusion method for generating protein structures with or without conditioning information such as a motif or target; it handles motif scaffolding, unconditional and symmetric (cyclic, dihedral and tetrahedral) generation, symmetric motif scaffolding, binder design and design diversification through partial diffusion.
• Nucleotide Transformer various · Science/Biology · 2017 — InstaDeep's Nucleotide Transformer, a family of genomics foundation language models developed with NVIDIA and TUM and pretrained on DNA sequences from over 3,200 diverse human genomes and 850 genomes from a wide range of species, in variants up to 2.5B parameters plus the later NT-v2 series; InstaDeep reports highly accurate molecular-phenotype prediction, published the work in Nature Methods in 2025 and has extended it with AgroNT, SegmentNT and NTv3.
• v0 (Vercel) Vercel · Code AI · 2023 — v0, Vercel's AI app builder: users prompt it to generate working full-stack web apps, which it plans, splits into tasks and connects to databases as it builds, then publish them as live sites with one-click deployment to Vercel; it syncs with GitHub and has a visual design mode, templates, design systems, an iOS app and v0 Mini, Pro and Max models. Free use is limited to 7 messages a day; Plus costs $30 and Business $100 per user a month.
• Viggle AI Viggle · Video Generation · 2024 — Character animation from text/image.
• Moondream 2 Vikhyat · Multimodal LLM · 2024 — Moondream 2, a small open vision-language model designed to run efficiently anywhere and updated frequently (its June 21, 2025 release added grounded reasoning, sharper object detection, faster text generation and better UI understanding); Moondream now calls it its previous-generation model, after Moondream 3 (Preview).
• Voyage 3 Voyage AI · Embeddings · 2024 — High-performance text embeddings.
• Voyage AI voyage-3 Voyage AI · Embeddings · 2024 — Voyage AI text embedding model, now listed among Voyage's older models; Voyage's current lineup is the voyage-4 family.
• Warp AI Warp · Terminal AI · 2023 — AI-powered terminal with command suggestions.
• WellSaid Labs WellSaid Labs · Speech Synthesis · 2023 — Enterprise AI voice generation.
• Windsurf SWE-1 Windsurf · Code LLM · 2025 — Windsurf's first in-house model family, launched in May 2025 and built for the whole software engineering process rather than coding alone: SWE-1 (tool-call reasoning near Claude 3.5 Sonnet at a lower serving cost), SWE-1-lite (unlimited for all users) and SWE-1-mini (powers Windsurf Tab).
• Wix ADI Wix · Web Design · 2023 — AI website design from prompts.
• Wolfram Alpha Wolfram · Computational Knowledge · 2009 — Computational knowledge engine. Math, science, data.
• ChatSonic Writesonic · Chat Interface · 2023 — Writesonic's AI chatbot with web access.
• Writesonic Writesonic · AI Writing · 2023 — AI content generation platform.
• Aurora (xAI) xAI · Weather/Science · 2024 — xAI's image generation model, code-named Aurora (Dec 2024): an autoregressive mixture-of-experts network trained to predict the next token from interleaved text and image data; strong at photorealistic rendering and following text instructions, and can edit user-provided images. Launched in Grok on X.
• Grok (X Premium+) xAI · Chat Interface · 2024 — Grok on X, xAI's assistant built into the X app and website, where it draws on real-time posts; it first opened to a limited number of US users in November 2023, and X Premium+ subscribers got Grok 4 in July 2025. Since October 12, 2026, new subscribers get it through The X Subscription (Plus at $8 a month to Ultra at $200), whose paid plans include X Premium benefits and share one usage pool across the Grok apps, the Grok tab on X and Cursor; existing X Premium plans continue.
• Grok 1 xAI · Multimodal LLM · 2023 — xAI's Grok-1, the language model behind Grok, which xAI announced on November 3, 2023 as an AI modeled after The Hitchhiker's Guide to the Galaxy that answers with wit and a rebellious streak and draws on real-time knowledge from X; developed over four months, it scored 63.2% on HumanEval and 73% on MMLU. On March 17, 2024 xAI released the base model's weights and architecture, a 314B-parameter mixture of experts with 25% of weights active per token, under Apache 2.0.
• Grok 2 xAI · Multimodal LLM · 2024 — xAI's Grok-2 (August 13, 2024), released in beta on X with the smaller Grok-2 mini as a major step up from Grok-1.5 in chat, coding and reasoning; an early version tested on LMSYS Chatbot Arena as 'sus-column-r' outscored Claude 3.5 Sonnet and GPT-4 Turbo, and it scored 87.5% on MMLU and 56.0% on GPQA, understanding text and images with real-time information from X, with API access planned that month. xAI later published its weights on Hugging Face.
• Grok 3 xAI · Multimodal LLM · 2025 — xAI's Grok 3 (February 2025), trained on its Colossus supercluster with 10x the compute of previous state-of-the-art models and launched in beta with Grok 3 mini; its reasoning mode, Grok 3 (Think), scored 93.3% on AIME 2025 at its highest test-time compute and 84.6% on GPQA, Grok 3 reached 1402 Elo in Chatbot Arena, and it came with the DeepSearch agent for X Premium and Premium+ users. xAI retired grok-3 from its API on May 15, 2026, redirecting it to Grok 4.3.
• Grok 4 xAI · Multimodal LLM · 2025 — Grok 4 (July 9, 2025), which xAI called the most intelligent model in the world: trained with RL at pretraining scale on the 200,000-GPU Colossus cluster, with native tool use, real-time search across X and the web, and a 256K-token API context with image input. It set a closed-model record of 15.9% on ARC-AGI-2, and Grok 4 Heavy, using parallel test-time compute for the new SuperGrok Heavy tier, scored 50.7% on Humanity's Last Exam (text-only). Its API was retired May 15, 2026.
• Grok 4 Fast xAI · Multimodal LLM · 2025 — Grok 4 Fast (September 19, 2025), xAI's cost-efficient reasoning model built on lessons from Grok 4: one unified model for reasoning and non-reasoning modes with a 2M-token context, trained end to end with tool-use RL for agentic web and X search. xAI said it matched Grok 4 on benchmarks with 40% fewer thinking tokens, cutting the price of equal performance by 98%, scored 92.0% on AIME 2025 and took #1 in LMArena's Search Arena. Its API was retired May 15, 2026, with requests sent to Grok 4.3.
• Grok 4.1 xAI · Reasoning LLM · 2025 — xAI's Grok 4.1 (November 17, 2025), released to all users on grok.com, X and the iOS and Android apps after a two-week silent rollout in which users preferred it to the previous model 64.78% of the time; tuned with large-scale reinforcement learning for style, personality, emotional intelligence and helpfulness, its Thinking mode took #1 on LMArena's Text Arena at 1483 Elo, and its hallucination rate on real info-seeking prompts fell to 4.22% from Grok 4 Fast's 12.09%.
• Grok 4.1 Fast xAI · Reasoning LLM · 2025 — xAI's Grok 4.1 Fast (November 19, 2025), which xAI called its best tool-calling model, with a 2M-token context and long-horizon reinforcement learning for consistent multi-turn agent performance; it launched with the Agent Tools API (X and web search, remote code execution and more) for agents in fields such as customer support and finance, at $0.20/$0.50 per million tokens. xAI retired it from its API on May 15, 2026, redirecting requests to Grok 4.3.
• Grok 4.20 xAI · Reasoning LLM · 2026 — Grok 4.20, xAI's early-2026 model, which it describes as fast and strong at agentic tool calling, with the lowest hallucination rate on the market and strict prompt adherence; the API offers reasoning and non-reasoning versions plus Grok 4.20 Multi-agent, in which several agents work in parallel on deep research. All take text and image input with a 1M-token context at $1.25/$2.50 per million tokens ($0.20 cached). SpaceXSI now calls Grok 4.7 its flagship.
• Grok 4.3 xAI · Reasoning LLM · 2026 — Grok 4.3, the xAI API model that took over requests to Grok 4, Grok 4 Fast, Grok 4.1 Fast and Grok 3 when xAI retired them on May 15, 2026; xAI then called it the fastest, most intelligent model it had built, topping leaderboards in agentic tool calling and instruction following. It takes text and image input with a 1M-token context, supports function calling, structured outputs and reasoning effort from none to xhigh, and costs $1.25/$2.50 per million tokens.
• Grok 4.5 xAI · LLM · 2026 — Grok 4.5 (July 2026), SpaceXSI's smartest model for coding, agentic tasks and knowledge work at launch: trained alongside Cursor on tens of thousands of NVIDIA GB300 GPUs, with RL across hundreds of thousands of mostly software-engineering tasks, it scored 83.3% on Terminal-Bench 2.1 and 64.7% on SWE-Bench Pro while using about 4.2x fewer output tokens per task than Claude Opus 4.8. It runs at about 80 tokens a second, became Grok Build's default model and costs $2/$6 per million tokens.
• Grok 4.6 xAI · LLM · 2026 — Grok 4.6 (August 12, 2026), the update to Grok 4.5 focused on long-running agents and ambitious interactive and visual work: a longer supplemental training run with curated model-generated data and an improved optimizer, then agentic RL on knowledge work, coding, kernel optimization and computer-aided design. SpaceXSI says it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index and scores 69.9% on CursorBench 3.2; its API has a 500K-token context at $2/$6 per million tokens.
• Grok 4.7 xAI · LLM · 2026 — SpaceXAI's flagship for coding and knowledge work: a new, larger base model with a longer RL run weighted toward multi-hour tasks, better self-verification and long-context handling, and native understanding of the Grok Bot harness. Official results: CursorBench 4.0 46.3%, DeepSWE v1.1 71.0%, EEBench 64.0%.
• Grok Aurora xAI · Image Generation · 2024 — xAI's Aurora (December 9, 2024), the autoregressive mixture-of-experts image model behind Grok's image generation on X, trained to predict the next token from billions of interleaved text and image examples from the internet; it excels at photorealistic rendering and precise instruction following and can take inspiration from or edit user-provided images. It rolled out on X in select countries first, and xAI now points users to its newer Grok Imagine models.
• Grok Code Fast 1 xAI · Code LLM · 2025 — Speedy, economical reasoning model for agentic coding, built from scratch on a new lightweight architecture; strongest in TypeScript, Python, Java, Rust, C++ and Go; 256K context with visible reasoning traces. Launched free for a limited time in GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode and Windsurf.
• Grok Imagine Video 1.5 xAI · Video Generation · 2026 — Grok Imagine Video 1.5, SpaceXSI's image-to-video model, generally available in the xAI API since June 2026, with a Video 1.5 Fast version on grok.com/imagine and in the Grok apps: sound effects, ambience and dialogue are generated in the same pass, speech is clearer and better synced, and motion holds together better, while Video 1.5 Fast makes a 6-second 720p clip in about 25 seconds, down from 40+. A July 31 update added text-to-video, reference-to-video and native 1080p.
• Grok Voice Think Fast 2.0 xAI · Speech AI · 2026 — Grok Voice 2.0 (July 29, 2026), xAI's speech-to-speech model that reasons while it speaks: it scored 82.9% on Artificial Analysis's Speech-to-Speech Quality Index (GPT-Realtime-2.1: 79.1%) and 56.5% on τ-voice Bench, cut time to first audio to 0.70s from 1.25s, and with 60% fewer reasoning tokens usually fires tool calls before its first sentence ends. Starlink A/B tests lifted sales conversion and support containment; it cost $0.08 a minute and gave way to Grok Voice 3.0 on October 14, 2026.
• Grok-1 (open) xAI · LLM · 2024 — xAI's Grok-1 open-weights release: a 314B-parameter mixture-of-experts model with 8 experts (2 active per token), 64 layers, an 8,192-token context and a 131,072-token SentencePiece vocabulary, published with JAX example code under Apache 2.0.
• SuperGrok xAI · Chat Interface · 2025 — SuperGrok, the paid subscription for Grok on grok.com and the Grok iOS and Android apps: it raises usage limits across chat, Grok Imagine image and video generation, voice and Grok Build, drawing on one shared weekly usage pool since June 2026, and a SuperGrok Heavy tier added Grok 4 Heavy in July 2025. Since October 12, 2026, new customers get its benefits in The X Subscription's Premium plan ($30 a month, which also includes Cursor Pro and X Premium+); existing SuperGrok plans continue.
• Xiaohongshu dots.llm1 Xiaohongshu/RedNote · LLM · 2024 — Large mixture-of-experts model that activates 14B of its 142B total parameters, from Xiaohongshu (RedNote). MIT license.
• Xiaohongshu/RedNote dots.llm1 Xiaohongshu/RedNote · LLM · 2025 — Large mixture-of-experts model that activates 14B of its 142B total parameters, from Xiaohongshu (RedNote). MIT license.
• Xiaomi MiMo-7B-RL-0530 Xiaomi · LLM · 2025 — May 30, 2025 update of Xiaomi's 7B reasoning model MiMo-7B-RL, trained with a much larger SFT set (about 6M examples, up from 500K) and a longer RL window (48K) for stronger math and code. MIT license.
• YaLM 100B Yandex · LLM · 2022 — 100B-parameter GPT-like model trained for 65 days on 800 A100 GPUs over 1.7TB of English and Russian text; released free under Apache 2.0.
• YandexGPT Yandex · LLM · 2023 — YandexGPT, Yandex's large language model, launched in May 2023 in its Alice virtual assistant: it generates and processes text, performs tasks, suggests ideas and keeps conversation context, and Yandex used it to summarize articles in Yandex Browser, condense product reviews in Search and Market and draft product descriptions for Market sellers. YandexGPT 2 followed in September 2023, and in December 2023 Yandex opened the YandexGPT API in Yandex Cloud to businesses worldwide.
• YandexGPT 2 Yandex · LLM · 2023 — YandexGPT 2, the second version of Yandex's language model, released to users in Russia through the Alice assistant on Yandex Stations, TVs, Yandex Browser and Search; Yandex enlarged the model, improved training and added new prompt examples and synthetic data, and its AI trainers rated its answers better than the first version's in 67% of side-by-side comparisons. Its training data ran to March 2023, and businesses worldwide could use it through the YandexGPT API from December 2023.
• YandexGPT 3 Lite Yandex · LLM · 2024 — YandexGPT 3 Lite, the lighter version of Yandex's third-generation YandexGPT, offered through Yandex Cloud's Foundation Models API for speed-sensitive work such as chatbots, spell checking and data analysis, at 20 kopecks per 1,000 tokens. Yandex improved pretraining data selection and added curriculum learning, a better reward model and grouped query attention, and reports 6 points more correct answers on YaMMLU_ru than YandexGPT 2 Lite and better answers in 68% of side-by-side tests.
• YandexGPT 3 Pro Yandex · LLM · 2024 — YandexGPT 3 Pro, the most powerful model in Yandex's third YandexGPT generation, opened to YandexGPT API users in March 2024 alongside a nearly 50% API price cut; tuned for business tasks such as customer support, sales, legal and marketing work, it extracts data to JSON and can be fine-tuned in Yandex DataSphere, and Yandex reports it beat YandexGPT 2 in 67% and ChatGPT 3.5 in 64% of side-by-side tests. From April 2024 it powered Alice Pro, as YandexGPT 3 took over the Alice assistant.
• You.com You.com · Search AI · 2023 — AI search engine with chat mode.
• Zapier AI Zapier · Automation AI · 2023 — AI actions in workflow automation.
• Zendesk AI Zendesk · Support AI · 2023 — AI customer service automation.
• ChatGLM-6B Zhipu AI · LLM · 2023 — ChatGLM-6B, an open Chinese-English bilingual dialogue model based on the General Language Model (GLM) framework with 6.2 billion parameters, trained on about 1T tokens with supervised fine-tuning, feedback bootstrap and RLHF; it runs locally on consumer GPUs (6GB of GPU memory at INT4), is fully open for academic research and free for commercial use after registration, and was succeeded by ChatGLM2-6B.
• CogVideoX Zhipu AI · Video Generation · 2024 — Zhipu AI's CogVideoX open video generation models: CogVideoX-2B (August 2024, now Apache 2.0) and CogVideoX-5B cover text-to-video and video continuation, CogVideoX-5B-I2V adds image-to-video, and CogVideoX1.5-5B (November 2024) makes 10-second, higher-resolution videos; the 2B model runs on GPUs as old as a GTX 1080 Ti and the 5B on an RTX 3060.
• CogVLM Zhipu AI · Multimodal LLM · 2023 — CogVLM, an open visual language foundation model that adds a trainable visual expert module to the attention and FFN layers of a frozen pretrained language model for deep vision-language fusion without losing NLP performance; CogVLM-17B set state-of-the-art results on 10 cross-modal benchmarks and ranked second on VQAv2, OKVQA, TextVQA and COCO captioning, matching or beating PaLI-X 55B.
• CogVLM2 Zhipu AI · Multimodal LLM · 2024 — CogVLM2, a family of open visual language models for image and video understanding (CogVLM2, CogVLM2-Video and GLM-4V) that keeps CogVLM's visual-expert architecture with improved pre- and post-training, accepts images up to 1344x1344 pixels and adds video input with timestamps; its authors report state-of-the-art results on MMBench, MM-Vet, TextVQA, MVBench and VCGBench.
• GLM-4 Zhipu AI · Multimodal LLM · 2024 — Zhipu AI's GLM-4 model with a 128K context window; since succeeded by newer models, including GLM-4.6, GLM-4.7 and the GLM-5 series.
• GLM-4-0414 Zhipu AI · LLM · 2025 — Zhipu AI's GLM-4-0414 series of open models, released April 14, 2025: GLM-4-32B-0414, whose base model was pretrained on 15T tokens with much synthetic reasoning data; the deep-reasoning GLM-Z1-32B-0414; GLM-Z1-Rumination-32B-0414, which can use search tools while thinking; and the smaller GLM-4-9B-0414 and GLM-Z1-9B-0414, mostly trained at a native 32K context extendable to 128K. Zhipu says the 32B base model matches GPT-4o and DeepSeek-V3-0324 on some code and Q&A benchmarks.
• GLM-4.5 Zhipu AI · LLM · 2025 — Zhipu's July 2025 open-weight agentic flagship: 355B-total / 32B-active MoE (plus GLM-4.5-Air, 106B / 12B active) with hybrid thinking / non-thinking modes, 128K context, MIT license.
• GLM-4.7 Zhipu AI · LLM · 2025 — Z.ai's Dec 22, 2025 flagship (~358B-parameter MoE, MIT-licensed weights) built for agentic coding, tool use and multi-step reasoning, with interleaved and preserved thinking; Z.ai reports BrowseComp 67.5 and made it the default model for its GLM Coding Plan.
• GLM-5 Zhipu AI · LLM · 2026 — Z.ai's (formerly Zhipu AI) flagship MoE at its February 2026 launch (~744B params, 40B active, 200K context). First open-weights model to reach 50 on the Artificial Analysis Intelligence Index. Trained on 100,000 Huawei Ascend 910B chips - no Nvidia hardware. Weights on Hugging Face (zai-org/GLM-5). Z.ai now lists GLM-5.3 as its flagship.
• GLM-5.1 Zhipu AI · LLM · 2026 — Post-training upgrade to GLM-5 for agentic engineering (same ~744B MoE architecture, 40B active, 200K context, 131K max output). Topped SWE-Bench Pro at release (58.4) ahead of GPT-5.4 (57.7) and Claude Opus 4.6 (57.3) - first open-source model to lead that benchmark. Runs plan-execute-test-fix loops autonomously up to ~8 hours. Open weights released Apr 7, 2026.
• GLM-5.2 Zhipu AI · LLM · 2026 — 753B open MoE, MIT-licensed, 1M-token context — the open coding/agentic model that arrived exactly as Fable access paused. Later notable as the self-hosted model Hugging Face used for forensics on the July agent-intrusion incident (closed frontier models refused to help), and the focus model of AA's Endpoint Accuracy Index (the same weights score 52-100% depending on provider). 71% Terminal Bench 2.1 in REAP-pruned form.
• GLM-5.3 Zhipu AI · LLM · 2026 — Post-training-focused update on the GLM-5.2 base: same pretrained base, refined post-training, 1M-token context. $1.40/$4.40 per M via Z.ai.
• GLM-5.3-Flash Zhipu AI · Multimodal LLM · 2026 — Open multimodal MoE: 320B total / 18B active, MIT license, 1.05M-token context, very low pricing. Flash-tier sibling to GLM-5.3.
• Zhipu AI (Tsinghua) GLM-4-0414 Zhipu AI · LLM · 2025 — 32B GLM-4 update (Apr 2025) pre-trained on 15T tokens, including substantial reasoning-type synthetic data; Zhipu says it performs comparably to OpenAI's GPT series and DeepSeek V3/R1. MIT license.
• Zoom AI Companion Zoom · Meeting AI · 2023 — AI features in Zoom. Summaries, smart recording.