Changelog

1,176 models · 329 vendors · 17 fields corrected on 14 models

Corrected

• 14 entries checked — Fourteen entries checked against Databricks’, Upstage’s, Snowflake’s, Together’s, Arc Institute’s and OpenAI’s launch posts, the Swallow project’s page, Cerebras’ deprecation log and Hugging Face model cards; 17 fields corrected
• Descriptions (14) — MPT-30B, MPT-7B, both DBRX entries, SOLAR 10.7B, Jais 30B, Snowflake Arctic, RedPajama, Swallow, Cerebras Llama 3.3 70B, Arc Institute Evo 2, GPT-2 and both dots.llm1 entries now give official architectures, training data, results and licenses
• Errors fixed — Evo 2’s text was a placeholder that described nothing; GPT-2’s text called its withheld release the first time an AI lab self-censored a release, a claim not in OpenAI’s posts
• Primary sources (3) — RedPajama pointed to Together AI’s homepage, Swallow to the Swallow project’s homepage and MPT-7B to a general Databricks product page; each now points to its own page

1,176 models · 329 vendors · 17 fields corrected on 14 models

Corrected

• 14 entries checked — Fourteen entries checked against the original AlexNet, EVA, BloombergGPT and ULMFiT papers, Stanford’s GloVe page, Meta AI’s DINOv2 post, EleutherAI’s GPT-Neo repository, eight Hugging Face model cards and the Bamba launch post; 17 fields corrected
• Descriptions (14) — METAGENE-1, Baichuan-M1-14B, Bamba-9B, Salamandra, Sesame CSM, StarCoder 2, Falcon-40B, EVA-01, DINOv2, AlexNet, BloombergGPT, ULMFiT, GloVe and GPT-Neo now give official architectures, training data, results and licenses
• Errors fixed — METAGENE-1’s text was a placeholder that described nothing; Salamandra, Falcon, EVA-01, DINOv2 and AlexNet carried generic fragments that did not describe the models
• Primary sources (3) — DINOv2’s link returned Meta’s page-not-available error and now points to the post’s current address; EVA-01 pointed to BAAI’s homepage and AlexNet to Alex Krizhevsky’s personal page, and each now points to its paper

1,176 models · 329 vendors · 16 fields corrected on 13 models

Corrected

• 13 entries checked — Thirteen open-model entries checked against Z.ai, Tencent, Baidu, Qwen, Tongyi Lab and Sea AI Lab model cards, Zhipu AI’s ChatGLM report, the Qwen team’s QwQ-Max-Preview post and Berkeley’s Starling and Koala posts; 16 fields corrected
• Descriptions (13) — GLM-5.3-Flash, GLM-5.3, GLM-4, Hy-MT2-30B-A3B, Hy-MT2-1.8B, ERNIE-4.5-21B-A3B-Thinking, QwQ-32B, QwQ-Max-Preview, QwenLong-L1-32B, both Sailor2 entries, Starling-7B and Koala now give official architectures, training, languages and licenses
• Errors fixed — Both Hy-MT2 entries said 33 language pairs, while Tencent says the models translate among 33 languages, and gave an 8K context that is not in the model cards; ERNIE-4.5-21B-A3B-Thinking’s WAVE SUMMIT launch venue is not in Baidu’s model card; Starling-7B’s license now carries its condition against use to compete with OpenAI
• Primary sources (2) — GLM-5.3 and one Sailor2 entry pointed to company homepages; each now points to its model card

1,176 models · 329 vendors · 17 fields corrected on 16 models

Corrected

• 16 entries checked — Sixteen open-model entries checked against TII’s Falcon posts and model cards, Cohere’s model cards and Command R7B announcement, and IBM, NVIDIA, Ai2, BigScience and Microsoft model cards; 17 fields corrected
• Descriptions (16) — Falcon-H1, Falcon-Edge, Falcon 3, Command R+, Command R, Command R7B, Granite-3.2-8B-Instruct, Granite 3.1 8B, Llama-3.1-Nemotron-Ultra-253B, Llama-3.3-Nemotron-Super-49B, OLMo 2 32B, Llama-3.1-Tulu-3-405B, OLMo, BLOOM, BLOOM-176B and Phi-3 Medium now give official architectures, training, languages, context lengths and licenses
• Errors fixed — The three Falcon entries shared one generic fragment with no details of the models; BLOOM’s slogan (‘The people’s model’) and an unsourced researcher count are gone
• Primary sources (1) — OLMo pointed to Ai2’s general OLMo project page; it now points to the OLMo-7B model card

1,176 models · 329 vendors · 12 fields corrected on 7 models

Corrected

• 7 entries checked — Seven music, speech, image and language model entries checked against Suno’s announcements and release notes, ElevenLabs’ posts and models documentation, Ideogram’s model page and model card, and OpenAI’s gpt-oss-120b model card; 12 fields corrected
• Descriptions (7) — Suno v3, Suno v4, Suno v5.5, ElevenLabs v3, ElevenLabs Turbo v2, Ideogram 4.0 and OpenAI gpt-oss-120b now give official capabilities, features, licenses and availability; Suno v4’s ‘AI music leader’ and an editorial judgment on Ideogram 4.0 are gone
• Errors fixed — Suno v5.5 was compared with v4, while it followed Suno’s v5, and was said to add custom voice models, while Suno describes separate Voices and Custom Models features
• Primary sources (5) — Suno v3, Suno v4, ElevenLabs v3 and ElevenLabs Turbo v2 pointed to a company homepage and Ideogram 4.0 to an address that now redirects to Ideogram’s features page; each now points to its own announcement, model page or documentation

1,176 models · 329 vendors · 17 fields corrected on 12 models

Corrected

• 12 entries checked — Twelve Mistral AI entries checked against Mistral’s announcements, model documentation pages and deprecated-models table; 17 fields corrected
• Descriptions (12) — Robostral Navigate, Mistral Large 3, Ministral 3, Mistral Medium 3, Mistral Small 3, Mistral Small 3.1, Mistral Saba, Pixtral Large, Voxtral, Mistral OCR 3, Mistral Large (2402) and Mistral-7B-v0.2 now give official architectures, capabilities, benchmarks, prices, licenses and API retirement dates; Mistral Large 3’s ‘EU frontier anchor’ characterization is gone
• Errors fixed — Pixtral Large’s access terms said research and non-commercial use only, while Mistral also offers a Mistral Commercial License for commercial use
• Primary sources (4) — Mistral Medium 3, Mistral Large and Mistral-7B-v0.2 pointed to Mistral’s general models documentation and Mistral OCR 3 to its changelog; each now points to the model’s own announcement or documentation page

1,176 models · 329 vendors · 11 fields corrected on 7 models

Corrected

• 7 entries checked — Seven video and language model entries checked against Kuaishou’s press releases and quarterly results, Luma AI’s announcements, the KAT-V1 and Pyramid Flow papers and Hugging Face model cards; 11 fields corrected
• Descriptions (7) — Kling 1.5, Kling 2.5 Turbo, Kling 3.0, Luma Ray3, KAT-V1-40B, Klear-46B-A2.5B and Pyramid Flow now give official capabilities, specifications, benchmarks, prices and licenses
• Errors fixed — Kling 3.0’s native 4K at 60fps is not in Kuaishou’s announcement, and its third-party rankings are gone; Luma Ray3 tied a Draft Mode cost cut to Ray3.14, which Luma says is 3x cheaper at 720p than Ray3; Pyramid Flow gave only an MIT license and an unsourced ~16GB VRAM figure, while its weights are Apache 2.0 (miniFLUX) or Stability AI Community License (SD3)
• Primary sources (4) — Kling 1.5, Kling 3.0 and Luma Ray3 pointed to a homepage or general product page and Kling 2.5 Turbo to a news site’s press page; each now points to Kuaishou’s or Luma’s own release

1,176 models · 329 vendors · 10 fields corrected on 5 models

Corrected

• 5 entries checked — Five video and 3D generation entries checked against Kuaishou’s press releases and Luma AI’s launch posts and changelog; 10 fields corrected
• Descriptions (5) — Kling 1.0, Kling 2.0, Luma Dream Machine, Luma Genie and Luma Ray2 now give official architectures, capabilities, specifications and availability
• Errors fixed — Luma Genie was described as Dream Machine text-to-video, while Luma’s announcement calls it a generative 3D model; Kling 2.0’s 2-minute video length is not in Kuaishou’s Kling 2.0 announcement
• Primary sources (5) — All five pointed to the Kling AI or Luma AI homepage; each now points to its own press release or announcement

1,176 models · 329 vendors · 10 fields corrected on 6 models

Corrected

• 6 entries checked — Six Meta entries checked against Meta’s announcements, Meta developer model pages and blog, Hugging Face model cards and Hugging Face’s Muse Glimmer launch post; 10 fields corrected
• Descriptions (6) — Muse Spark 1.3, Muse Glimmer, Meta Muse Spark, Llama 3.3, Llama 3.2 90B Vision and Segment Anything Model (SAM) now give official capabilities, benchmarks, prices and licenses; the Muse Spark entry relied on third-party rankings, and Muse Spark 1.3’s contributor-tier price is now given as $0.10 input and $0.20 output
• Primary sources (4) — Muse Spark 1.3, Muse Glimmer, Llama 3.3 and Llama 3.2 90B Vision pointed to the AI at Meta homepage or Meta’s general Llama page; each now points to its own announcement, model page or model card

1,176 models · 329 vendors · 11 fields corrected on 7 models

Corrected

• 7 entries checked — Seven Midjourney entries checked against Midjourney’s Version and Legacy Features documentation and its update posts; 11 fields corrected
• Descriptions (5) — Midjourney v5, v5.2, v6.1, V8.1 and Midjourney Alpha (Web) now give release and default-model dates and official changes; V8.1 now notes that V8.2 replaced it as the default in July 2026, and its unsourced personalization-profile requirement is gone
• Primary sources (6) — Midjourney v5, v5.2, v6, v6.1, v7 and Midjourney Alpha (Web) pointed to Midjourney’s homepage; each now points to Midjourney’s documentation or the model’s own announcement

1,176 models · 329 vendors · 13 fields corrected on 7 models

Corrected

• 7 entries checked — Seven Google and Perplexity entries checked against Google’s and Perplexity’s announcements, help centers, API documentation and pricing page; 13 fields corrected
• Descriptions (7) — Google Gemini for Workspace, Google Assistant (Gemini), Socratic (Google), Photomath, Sonar Large, Perplexity Sonar and Perplexity Pro now give official features, history, prices and current status
• Primary sources (6) — Gemini for Workspace, Google Assistant (Gemini) and Socratic pointed to Gemini API developer pages, and Sonar Large, Perplexity Sonar and Perplexity Pro to Perplexity’s homepage or documentation root; each now points to its own announcement, help article or changelog

1,176 models · 329 vendors · 9 fields corrected on 8 models

Corrected

• 8 entries checked — Eight Google and Google DeepMind research entries checked against their papers, Google and Google DeepMind announcements, the Nobel Prize press release and a Hugging Face model card; 9 fields corrected
• Descriptions (8) — Transformer, BERT, Chinchilla, AlphaFold 2, Gemma 3 27B, Gemini Robotics, Gemini Robotics-ER and Gemini Robotics On-Device now give official capabilities, results and availability
• Errors fixed — Gemini Robotics On-Device, which runs locally on robots, was listed as cloud only; the BERT, Chinchilla and AlphaFold 2 entries were short fragments with unsourced claims

1,176 models · 329 vendors · 20 fields corrected on 11 models

Corrected

• 11 entries checked — Eleven Google entries checked against Google’s announcements, Google DeepMind’s model cards and model pages, and the Gemini API deprecations page; 20 fields corrected
• Descriptions (11) — Gemini 3.8 Flash, Gemini 3.8 Flash Cyber, Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, Gemini 3 Deep Think, Gemini Robotics-ER 1.6, Gemma 4, Nano Banana, Gemini 2.5 Pro 06-05, Gemini 2.5 Flash Preview and Gemini Flash now give official capabilities, benchmarks, prices and current status
• Errors fixed — Gemini 3.8 Flash’s Terminal-Bench 2.1 score was given as 90.8%, while Google’s model card reports 89.4%; Gemini 3.5 Flash’s ‘25-40% lower cost’ and Gemini 3.1 Flash-Lite’s ‘default gateway model’ claims are not in Google’s sources; the Gemini 3 Deep Think entry described its place in this database rather than the model; Gemini Robotics-ER 1.6 was listed as on-device, while Google offered it through the Gemini API and Google AI Studio
• Primary sources (8) — Gemini 3.8 Flash Cyber, Gemini 3.5 Flash, Gemini 3.1 Flash-Lite, Gemini 3 Deep Think, Gemini Robotics-ER 1.6, Gemini 2.5 Pro 06-05, Gemini 2.5 Flash Preview and Gemini Flash now point to their own announcements or model pages instead of general Gemini pages

1,176 models · 329 vendors · 10 fields corrected on 8 models

Corrected

• 8 entries checked — Eight OpenAI entries checked against OpenAI’s announcements, API model and deprecation pages, papers, Hugging Face cards and GitHub repositories; 10 fields corrected
• Descriptions (8) — GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, gpt-oss-20b, OpenAI o1, o1-2024-12-17, CLIP (ViT L/14), GPT-3 and GPT-1 now give official capabilities, benchmarks, prices and current status
• Errors fixed — The o1-2024-12-17 entry described the original o1 and dated it September 2024, while the snapshot came out in December 2024; GPT-1’s 117M parameter count is not in OpenAI’s paper; o1 and o1-2024-12-17 now note their October 23, 2026 API shutdown, and GPT-3 its January 4, 2024 retirement
• Primary sources (2) — GPT Image 2.5 Flare and Sunburst now point to OpenAI’s ChatGPT Images 2.5 announcement instead of OpenAI’s general news index

1,176 models · 329 vendors · 16 fields corrected on 12 models

Corrected

• 12 entries checked — Twelve Microsoft entries checked against Microsoft Learn, Microsoft Support, the Bing and Microsoft AI blogs, LinkedIn Help and the WizardLM project page; 16 fields corrected
• Descriptions (12) — Copilot (consumer), Copilot App, Bing Chat / Copilot, Bing Copilot, Microsoft Copilot for M365, Microsoft Security Copilot, LinkedIn AI Writing, WizardLM 2, MAI-Code-1-Flash, MAI-Image-2.5, MAI-Transcribe-1.5 and MAI-Voice-2
• Errors fixed — LinkedIn’s AI writing tool is no longer available, per LinkedIn’s help page; Copilot (consumer)’s entry named GPT-4 as its model, which Microsoft’s current pages do not; MAI-Transcribe-1.5 ranks third on Artificial Analysis, not first; MAI-Image-2.5’s lower-price claim is not in Microsoft AI’s announcement
• Deployment (1) — MAI-Image-2.5 now lists Foundry, OpenRouter, PowerPoint and OneDrive
• Primary sources (3) — Microsoft Copilot for M365 now points to Microsoft’s Copilot overview instead of Microsoft’s general AI page, LinkedIn AI Writing to its help article instead of LinkedIn’s homepage, and MAI-Voice-2 to its announcement’s current address

1,176 models · 329 vendors · 18 fields corrected on 8 models

Corrected

• 8 entries checked — Eight AI coding tools checked against their makers’ websites, pricing pages, docs and announcements; 18 fields corrected
• Descriptions (8) — Cursor, GitHub Copilot, Codeium, Devin, Lovable.dev, Replit Agent, bolt.new and v0 (Vercel) now give official capabilities, prices and current status
• Status changes — SpaceX acquired Cursor in August 2026, and Cursor is now sold through The X Subscription; Codeium renamed itself Windsurf in April 2025, Cognition agreed to buy it in July 2025, and the Windsurf IDE is now Devin Desktop
• Prices (2) — Cursor’s entry still listed its old $20 Pro and $40 Business plans; GitHub Copilot’s listed an ‘Individual’ plan and left out Pro+, Max and Enterprise
• Free tier (1) — GitHub Copilot Free is open to individual developers in general, not only students
• Deployment (2) — Cursor and GitHub Copilot now include their CLIs, apps and cloud agents
• Primary sources (4) — Codeium, Replit Agent and v0 now point to the pages their old addresses redirect to (Devin Desktop, Replit’s Agent page and v0.app), and Devin to devin.ai instead of Cognition’s company homepage
• X handle (1) — Codeium’s blank handle filled with @cognition, the account Devin Desktop’s page links

1,176 models · 329 vendors · 40 fields corrected on 20 models

Corrected

• 20 entries checked — All 20 Grok entries checked against SpaceXSI’s (xAI’s) announcements, API docs, release notes and plan pages; 40 fields corrected
• Descriptions (10) — Grok 4.6, Grok 4.5, Grok 4.20, Grok 4.3, Grok 4 Fast, Grok 4, Grok Voice Think Fast 2.0, Grok Imagine Video 1.5, SuperGrok and Grok (X Premium+) now give official capabilities, benchmarks, prices and current status
• Errors fixed — Grok 4.20’s entry gave a 2M-token context (xAI’s model page lists 1M); Grok 4.5’s 1.5-trillion-parameter, ‘V9 base’ and Cursor-data claims and Grok 4.6’s ‘V9 base’ are not in SpaceXSI’s announcements; Grok Imagine Video 1.5’s #1-leaderboard claim is not in its announcement; Grok Voice Think Fast 2.0 (officially Grok Voice 2.0) was A/B-tested on Starlink’s phone line, not already in production there, and has been deprecated for Grok Voice 3.0
• Prices (2) — SuperGrok’s entry gave the X Premium+ price; the SuperGrok and Grok (X Premium+) prices now reflect The X Subscription, which replaced SuperGrok and X Premium for new customers on October 12, 2026
• Deployment (3) — Grok 4.6, Grok 4.5 and Grok Voice Think Fast 2.0 now list the channels SpaceXSI names
• Primary sources (5) — Grok 4.3 and Grok 4.20 now point to their model pages, Grok 4.5 to SpaceXSI’s announcement instead of Cursor’s post, Grok Imagine Video 1.5 to its news address and SuperGrok to The X Subscription announcement instead of the company homepage
• X handle (20) — @SpaceXAI is now @SpaceXSI on every xAI entry, as linked from SpaceXSI’s website

1,176 models · 329 vendors · 15 fields corrected on 9 models

Corrected

• 9 entries checked — Seven descriptions that ran past the 500-character limit, from Claude Fable 5 and Claude Mythos 5 to GPT-5.6 Sol and FLUX 3, plus Hunyuan-A13B and YOLO-World, checked against official announcements, docs, model cards and papers; 15 fields corrected
• Descriptions (9) — Claude Fable 5, Claude Mythos 5, Claude Mythos Preview, GPT-5.6 Sol, GPT-Live, FLUX 3, WildGuard, Hunyuan-A13B and YOLO-World now fit the limit and give official capabilities, benchmarks, prices and current status
• Errors fixed — GPT-5.6 Sol’s entry claimed a customer-by-customer Commerce Department review and a roughly 4T-parameter ‘Spud’ pretrain, which OpenAI’s pages do not say; GPT-Live’s wake-word and desktop claims are not in OpenAI’s announcement; Claude Mythos 5’s entry said access was restored globally on July 1 (Anthropic restored it only for approved US organizations); Claude Mythos Preview’s entry did not note that Mythos 5 replaced it (Anthropic lists it as deprecated); Fable 5 and Mythos 5 now note their successors, Fable 5.1 and Mythos 5.1
• Prices (2) — Claude Fable 5 moved to usage credits in Claude plans after July 7, 2026, not June 22; GPT-Live-1’s API price ($0.05 a minute) added
• Deployment (3) — GPT-5.6 Sol (ChatGPT, Codex and the API, plus Cerebras for select customers); GPT-Live (now also in the API); FLUX 3 (BFL API and select partners)
• Primary sources (1) — YOLO-World now points to its GitHub repository instead of Tencent AI Lab’s homepage

1,176 models · 329 vendors · 23 fields corrected on 15 models

Corrected

• 15 entries checked — Short, outdated or over-long descriptions for Qwen, DeepSeek and Kimi models, from DeepSeek LLM and DeepSeek-R1 to Qwen3.8-Max-0902 and Kimi K3, checked against official model cards, release notes and Qwen Cloud and Model Studio pages; 23 fields corrected
• Descriptions (15) — Qwen3.8-Max, Qwen3.8-Max-0902, Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen3.7-Max and Wan 3.0; DeepSeek-R1, DeepSeek-V3, DeepSeek LLM, DeepSeek-V4.1-Flash and DeepSeek-V4-Flash-Vision-Exp; Kimi K3, Kimi K2.5, Kimi K2 Thinking and Kimi now give sizes, benchmarks, prices and current status
• Errors fixed — Qwen3.8-Max and Kimi K3 carried stock-market moves and third-party rankings not found in the vendors’ materials, and Kimi K3 and DeepSeek-V4.1-Flash ran past the 500-character limit; Qwen3.7-Max was called the first closed-weight Qwen flagship; Wan 3.0 was still described as in public beta (generally available since August 24, 2026); Qwen3.8-27B was called weights-only though Qwen Cloud hosts it; DeepSeek-V4-Flash-Vision-Exp’s image pricing was misdescribed; Kimi K2.5’s size was given as 1.04T parameters (Moonshot lists 1T)
• Prices (1) — Qwen3.8-Flash-Next’s API price is for the hosted Qwen3.8-Flash
• Deployment (2) — Wan 3.0 is generally available on Qwen Cloud and Model Studio; Kimi K3’s deployment field held dates and a license judgment
• Primary sources (5) — Qwen3.8-Max-0902, Qwen3.8-27B and Qwen3.8-Flash-Next pointed to the index of Qwen’s blog; DeepSeek-V4-Flash-Vision-Exp to the front page of DeepSeek’s API documentation; Kimi to Moonshot’s Chinese corporate homepage

1,176 models · 329 vendors · 28 fields corrected on 16 models

Corrected

• 16 entries checked — Short or outdated descriptions for Inflection AI, MiniMax and Qwen models, from Pi and Inflection-2.5 to MiniMax H3 and Qwen3-Coder, checked against official announcements, model cards, licenses and help pages; 28 fields corrected
• Descriptions (16) — Inflection-2.5, Inflection-3.0, Pi and Inflection Pi; MiniMax-M2.7, M2.5, M2, Text-01, M3, H3, Hailuo AI and Talkie; Qwen3-Coder, Qwen3, Wan and Qwen3.6-35B-A3B now give sizes, benchmarks, features and licenses
• Errors fixed — MiniMax-M2.7 was described as natively multimodal with voice, image and video, which MiniMax does not say; MiniMax H3’s license was called litigation-linked (it excludes the US, EU, UK and South Korea); MiniMax-M3 still said its weights were pending; MiniMax-M2.5’s license was listed as MIT (it is modified MIT); Qwen3 was called competitive with GPT-4o and ended in an unsourced judgment
• Prices (4) — Pi and Inflection Pi are free consumer apps, not API products; Talkie is free with an optional Talkie+ upgrade; MiniMax-M3’s license allows commercial use with conditions, not only non-commercial use
• Deployment (2) — Qwen3-Coder also runs on Alibaba Cloud Model Studio’s API, and MiniMax H3 on MiniMax’s API and Hailuo AI
• Primary sources (6) — Inflection-3.0, Pi and Inflection Pi pointed to Inflection’s homepage; Hailuo AI and Talkie to MiniMax’s Chinese homepage; Qwen3-Coder to the general Qwen3 announcement

1,176 models · 329 vendors · 24 fields corrected on 20 models

Corrected

• 20 entries checked — Short descriptions for open-source community tools, InternLM-family models and YandexGPT, from AudioLDM 2 and Fooocus to Intern-S1 and YandexGPT 3 Lite, checked against official repositories, model cards, papers and announcements; 24 fields corrected
• Descriptions (20) — AudioLDM 2, Aider, ComfyUI, Continue, DiffusionBee, Draw Things, Flowise, Fooocus, GeoGPT, Langflow, Open WebUI and Piper; OREAL-32B, Intern-S1, InternLM3 and InternVL 2.5; YandexGPT, YandexGPT 2, YandexGPT 3 Pro and YandexGPT 3 Lite now give capabilities, sizes, benchmarks and current status
• Errors fixed — ComfyUI was described only as a Stable Diffusion interface (it now runs image, video, 3D and audio models); Continue and Flowise were listed without noting that Continue’s repository is read-only after its acquisition by Cursor and that Flowise was wound down in August 2026; Fooocus’s limited long-term support is now noted; the four InternLM-family rows only repeated the vendor and model name; the four YandexGPT rows shared one generic text ending in an unsourced judgment
• Primary sources (4) — Piper’s old repository now points to the Open Home Foundation’s piper1-gpl; YandexGPT 2, 3 Pro and 3 Lite pointed to Yandex Cloud’s general foundation models page

1,176 models · 329 vendors · 31 fields corrected on 21 models

Corrected

• 21 entries checked — Short Runway, Amazon, Black Forest Labs and Adobe descriptions, from Stable Diffusion and CodeWhisperer to Gen-4, Alexa+ and FLUX1.1 [pro] Ultra, checked against official announcements, product pages, documentation and model cards; 31 fields corrected
• Descriptions (21) — Runway Gen-1, Gen-2, Gen-3 Alpha, Gen-3 Alpha Turbo and Gen-4, the Runway platform and Stable Diffusion (LDM-KL-8-G); Alexa+, Amazon Nova, Nova Pro, Amazon Q Developer, CodeWhisperer and Titan Text; FLUX 1.1 Pro, FLUX 1.1 Ultra, FLUX.1 Schnell and FLUX.1 Dev; Adobe Firefly 1, 2 and 3 and Adobe Podcast now give capabilities, sizes, prices and retirement dates
• Errors fixed — Runway Gen-4 was credited with lip sync, generative audio and frame interpolation, which are not on Runway’s Gen-4 page; Amazon Nova and Nova Pro shared the same generic text; retirements are now noted for Runway Gen-1 (April 6, 2025), Gen-2 (May 11, 2025), Gen-3 Alpha (July 8, 2026) and Gen-3 Alpha Turbo (July 30, 2026), along with CodeWhisperer’s move into Amazon Q Developer and Titan Text’s absence from Bedrock’s current Titan list
• Primary sources (10) — Runway Gen-1, Gen-2, Gen-3 Alpha and Gen-3 Alpha Turbo pointed to Runway’s research hub; FLUX 1.1 Pro and Ultra to Black Forest Labs’ list of all models; Firefly 1, 2 and 3 to Adobe’s general Firefly page; CodeWhisperer to the Amazon Q Developer product page

1,176 models · 329 vendors · 30 fields corrected on 25 models

Corrected

• 25 entries checked — Short Meta and Stability AI descriptions, from OPT and Stable Diffusion 2.0 to Llama 4, SAM 3D and Muse Video, checked against official announcements, papers and model cards; 30 fields corrected
• Descriptions (25) — Muse Video, SAM 3D, SAM 2, Llama 4, Llama 4 Scout, Llama 3, LLaMA, Llama-3-Swallow-70B, Movie Gen, Chameleon, Meta AI, Voicebox, ImageBind, Galactica, OPT, Stable Diffusion 3.5, 3, 2.1 and 2.0, SDXL 1.0, Stable Video Diffusion, SV3D, Stable Audio, Stable Audio 2.0 and Clipdrop now give sizes, training data, release terms and reported results
• Errors fixed — Llama 4 Scout was given a 128K context window (Meta says 10 million tokens); Movie Gen was said to make 30-second videos (up to 16 seconds); SAM 3D was called 3D scene segmentation (it reconstructs 3D objects and people from one image); Llama 3’s license was called “Apache-like” (it is Meta’s Llama 3 Community License); LLaMA was called open source (Meta released it case by case under a noncommercial research license); claims that LLaMA leaked and that Galactica was pulled offline after three days are not in Meta’s sources; unsourced judgments were removed from Llama 4, Llama 3 and OPT
• Primary sources (5) — Llama 3, LLaMA and Llama-3-Swallow-70B pointed to a generic Llama link that now redirects to Meta’s developer homepage; Stable Diffusion 2.0 and 2.1 pointed to Stability AI’s homepage

1,176 models · 329 vendors · 23 fields corrected on 23 models

Corrected

• 23 entries checked — Short Google and DeepMind research descriptions, from word2vec and the 2014 sequence-to-sequence LSTM to PaLM 2, Gato and AlphaFold 3, checked against their papers and official research posts; 23 fields corrected
• Descriptions (23) — PaLM (two rows), PaLM 2, PaLM-E, FLAN, Minerva, Gato, GLaM, Gopher, CoAtNet, LaMDA, EfficientNetV2-XL, ViT-Base, MuZero, T5, T5-11B, XLNet, BERT-Large, the 2014 sequence-to-sequence LSTM, word2vec, AlphaGo, AlphaFold 3 and MusicLM now give sizes, training data, methods and benchmark results
• Errors fixed — FLAN was called an instruction-tuned PaLM (its paper tuned LaMDA-PT, a 137B model); the 2014 sequence-to-sequence model was said to use attention (it encodes each sentence into one fixed-size vector); Minerva was credited with “PhD-level math” (its paper reports undergraduate-level results); LaMDA’s entry repeated a sentience claim that is not from an official Google source; T5’s entry had a stray quotation mark; word2vec and PaLM entries carried unsourced judgments

1,176 models · 329 vendors · 38 fields corrected on 22 models

Corrected

• 22 entries checked — Short Google descriptions, from Bard and Google Translate’s neural system to Gemini 3.7 Flash, Nano Banana 2 Lite and Lyria 3, checked against Google’s announcements, model cards and developer docs; 38 fields corrected
• Descriptions (22) — Gemini 3.7 Flash, Gemini 3.5 Flash-Lite, Gemini 3.5 Flash Cyber, Gemini 3 Flash, Gemini Deep Think, Gemini CLI, Gemini Canvas, the Gemini app, Gemini Nano, Nano Banana 2 Lite, Nano Banana Pro, Imagen 3, Imagen 4, Veo 2, Veo 3, Veo 3.1, Lyria 2, Lyria 3, Flow, AI Overviews, Bard and Google Translate’s neural system now give release details, capabilities, benchmark results, availability and prices
• Errors fixed — Lyria 3 claimed “70+ languages” (Google lists eight for the Gemini app at launch); Veo 3 was called a “Market leader”; Nano Banana 2 Lite was sourced from a third-party newsletter and gave a sub-4-second time where Google says 4 seconds; Lyria 2 was said to power MusicFX (Google credits MusicFX DJ to Lyria RealTime)
• Primary sources (15) — Rows that pointed at Google DeepMind’s pages for its current Gemini, Veo, Lyria and Imagen models, at the Gemini API model list or at a newsletter now point at their own announcements, model cards or papers
• Subscription cost (1) — Gemini 3.5 Flash Cyber now notes its limited pilot for governments and trusted partners

1,176 models · 329 vendors · 40 fields corrected on 26 models

Corrected

• 26 entries checked — Short Google descriptions, from the first Gemma and Gemini 1.0 to Gemini 3 Pro, TranslateGemma and EmbeddingGemma, checked against Google’s model cards, technical reports, Gemma release log and launch posts; 40 fields corrected
• Descriptions (26) — Gemma (two rows), Gemma 3, Gemma 3 270M, Gemma 3n, CodeGemma, PaliGemma 2, MedGemma, TxGemma, Agentic-Tx, VaultGemma, T5Gemma, TranslateGemma, EmbeddingGemma, Gemini 1.0, Gemini Ultra (three rows), Gemini 2.0, Gemini 2.0 Flash (experimental), Gemini 2.0 Pro, Gemini 2.5, Gemini 2.5 Pro, Gemini 2.5 Flash, Gemini 3 and Gemini 3 Pro now give release details, sizes, context windows, benchmark results and availability
• Errors fixed — Eight Gemma-family rows shared one generic sentence (“Google’s open-weight gift to the community. Small, efficient, surprisingly capable for its size.”); three Gemini 2.0 rows, two Gemini 2.5 rows and two Gemini Ultra rows each shared one fragment; Gemini 1.0 opened with “Google finally entered the chat”; Gemini 3’s 95% AIME 2025 figure, which could not be checked in the text of Google’s announcement, was dropped
• Primary sources (14) — Gemma rows that pointed at the Gemma release log, Gemini rows that pointed at Google DeepMind’s page for its current Gemini models, and Gemini 2.5 Flash, which pointed at the Gemini 2.5 Pro announcement, now point at their own announcements or papers

1,176 models · 329 vendors · 29 fields corrected on 20 models

Corrected

• 20 entries checked — Short Microsoft descriptions, from ResNet-50 to Phi-4-mini-flash-reasoning and the new MAI image and voice models, checked against Microsoft’s model cards, research papers and Microsoft AI announcements; 29 fields corrected
• Descriptions (20) — Phi-2, Phi-3, Phi-4, Phi-4-mini, Phi-4-multimodal, Phi-4-reasoning-plus, Phi-4-mini-flash-reasoning, Orca, Orca 2, Florence-2, E5, E5-Mistral-7B, BioGPT, LayoutLMv3, DeBERTaV3, PubMedBERT (now BiomedBERT), ResNet-50, MAI-Image-2.5-Pro and MAI-Voice-2-Flash now give sizes, training data, context windows, benchmark results, licenses and prices
• Errors fixed — Six Phi rows shared one generic sentence (“Open-weight. Strong math and reasoning for its size. Runs locally.”); Phi-2 was said to compete with 13B models (Microsoft compares it with models under 13B); DeBERTaV3 was called a “SuperGLUE leader”, which its paper does not claim
• Primary sources (9) — Research models that pointed at general Microsoft Research pages now point at their own papers

1,176 models · 329 vendors · 38 fields corrected on 23 models

Corrected

• 23 entries checked — Short Mistral AI and xAI descriptions, from Mistral 7B to Devstral 2 and from Grok-1 to Grok 4.1 Fast, checked against the companies’ announcements, model pages, retirement notices and Hugging Face model cards; 38 fields corrected
• Descriptions (23) — Mistral 7B, Mistral-7B-Instruct-v0.3, Mixtral 8x7B and 8x7B Instruct, Mixtral 8x22B, Mistral Embed, Le Chat, Codestral, Codestral Mamba, Mathstral, Mistral NeMo, Mistral Large 2, Pixtral 12B, Magistral Medium, Devstral Medium, Devstral 2, Grok-1, Grok-2, Aurora, Grok 3, Grok 4.1 and Grok 4.1 Fast now give launch dates, sizes, context windows, benchmark results and API retirement dates
• Errors fixed — Three descriptions were internal editing notes (Devstral 2, Devstral Medium, Grok 4.1 Fast); a Devstral Medium row gave it Devstral Small’s 24B parameters; Mixtral 8x7B was described as 8 experts of 7B each (46.7B total, 12.9B per token); Grok-2 named both Aurora and FLUX as its image generator; unsourced lines such as “France shot across the bow” are gone; Le Chat’s entry now notes that it became Vibe on May 28, 2026
• License (1) — Grok 2’s open weights now carry their exact license name, the Grok 2 Community License Agreement
• Primary sources (14) — Mistral rows that pointed at the overview of Mistral’s models now point at each model’s announcement or model card

1,176 models · 329 vendors · 13 fields corrected on 12 models

Corrected

• 12 entries checked — The last OpenAI rows with fragment descriptions, from GPT-2 to GPT-5.6 Terra, checked against OpenAI’s announcements, model pages, deprecations page, API changelog and research papers; 13 fields corrected
• Descriptions (12) — GPT-2, GPT-3, InstructGPT, DALL·E 2, GPT-4o, GPT-4.5, GPT-4.1, GPT-5.1, GPT-5.2, GPT-5.5 Pro, GPT-5.6 Terra and GPT-Live-Transcribe now give launch dates, context windows, prices, benchmark results and shutdown dates
• Errors fixed — GPT-5.5 Pro’s description was an internal editing note (“keep separate only if product/API copy requires it”); GPT-Live-Transcribe’s “~18% WER improvement” is not in OpenAI’s documentation; unsourced lines such as “The image generator that made the world pay attention” (DALL·E 2) are gone
• Primary source (1) — GPT-Live-Transcribe pointed at OpenAI’s general speech-to-text guide and now points at its own model page

1,176 models · 329 vendors · 30 fields corrected on 24 models

Corrected

• 24 entries checked — Every short OpenAI description we flagged, from GPT-3.5 Turbo to GPT-5.6 Luna, checked against OpenAI’s announcements, model pages, deprecations page, release notes and help center; 30 fields corrected
• Descriptions (24) — ChatGPT, GPT-3.5 Turbo, GPT-4, GPT-4 with vision, GPT-4 Turbo, GPT-4o mini, GPT-4.1 mini and nano, GPT-5.4, GPT-5.6 Luna, o1-preview, o1-mini, o3-pro, gpt-oss-120b and gpt-oss-20b, DALL·E 3, GPT Image 1, Whisper, CLIP, the three text embedding models, ChatGPT canvas and ChatGPT agent now give launch dates, context windows, prices, benchmark results and shutdown dates
• Errors fixed — GPT-4 was listed with a 128K-token context (it shipped with 8,192 and 32,768 tokens; 128K came with GPT-4 Turbo); gpt-oss was called OpenAI’s first open-source model (OpenAI open-sourced Whisper in 2022); o1-preview was credited with PhD-level science accuracy, a claim OpenAI made about the update that followed it; ChatGPT’s “300M+ weekly active users” is now OpenAI’s August 2026 figure of more than 1 billion
• Status updates — OpenAI says ChatGPT agent is no longer available and points users to ChatGPT Work; canvas has left current ChatGPT models and stays only in legacy models for paid users, so its free tier is now No
• Primary sources (3) — ChatGPT agent, canvas and GPT-4 with vision pointed at the overview of OpenAI’s API models and now point at their own announcement or system card

1,176 models · 329 vendors · 44 fields corrected on 27 models

Corrected

• 27 entries checked — Every short Claude description, from Claude 1 to Sonnet 4.6, checked against Anthropic’s announcements, model deprecation page, pricing page and product pages; 44 fields corrected
• Descriptions (27) — Claude, Claude Instant, the Claude app, Claude 2 and 2.1, Claude 3 (family, Opus, Sonnet and Haiku), Claude 3.5 (family, Sonnet from June and October 2024, and Haiku), Projects, Claude 3.7 and 3.7 Sonnet, Claude 4, Opus 4 and Sonnet 4, Claude Code, Opus 4.1, Claude in Chrome, Sonnet 4.5, Haiku 4.5, Opus 4.5, Opus 4.6 and Sonnet 4.6 now give launch dates, context windows, prices, benchmark results and retirement dates
• Errors fixed — Claude 2 was listed with a 200K-token context (it launched with 100K; 200K came with Claude 2.1); the Claude 3.5 family row credited the June 2024 launch with computer use (added in October 2024); Opus 4.6’s 1M-token context was called generally available (it launched in beta); unsourced lines such as “The value play that won developers” are gone
• Primary sources (17) — Rows that pointed at the overview of Anthropic’s current API models (or the Opus line page) now point at each model’s or product’s own announcement or page

1,176 models · 329 vendors · 55 fields corrected on 37 models

Corrected

• 37 entries checked — More short descriptions and weak source links checked against technical reports, model cards, product documentation, announcements and READMEs; 55 fields corrected
• Descriptions (36) — Google’s Gemma 2 (family, 2B, 9B and 27B), ShieldGemma, Genie 2, AlphaZero, Gemini 1.5 (family, Pro and Flash) and Gemini 3.1 Pro; Jina Embeddings v3 and Jina Reranker v2; Cohere Embed v3 and Rerank 3; Vicuna; Dolly 2.0; Fuyu-8B; EVA-02; BEiT; Emu Edit; Make-A-Video; Llama Guard and Llama Guard 2; ModernBERT; IP-Adapter; ChemBERTa; StyleTTS 2; 3D Gaussian Splatting; Intel’s Neural Chat; Amazon Nova Lite and Micro, Titan Image Generator and Titan Embeddings; and SEA-LION now give sizes, training data, context lengths, licenses and results from the developers’ documentation, replacing fragments and an unsourced “1,500-page book” claim for Gemini 1.5
• Primary sources (18) — Links that pointed at a vendor homepage or a general list (Google’s Gemma release log, Google DeepMind’s Gemini page, the Gemini API model list, Jina AI’s model list, Meta’s Llama page, Intel’s AI page, and the BAAI, Microsoft Research, Answer.AI, Tencent AI Lab and DeepChem homepages) now point at each model’s own card, paper, project page or announcement
• Pricing (1) — SEA-LION was listed as MIT-licensed; its README says the terms vary with the Llama, Gemma or other base model

1,176 models · 329 vendors · 33 fields corrected on 29 models

Corrected

• 29 entries checked — More short descriptions checked against technical reports, model cards, project pages and READMEs; all 29 rewritten
• Descriptions (29) — Google DeepMind’s Gemma 3 1B, 4B and 12B and Gemma 2B and 7B; Cohere’s Aya 23, Aya Expanse and March 2024 Command R; Ai2’s WildGuard; Stable Audio Open; Med-PaLM and Med-PaLM 2; VGG16 and VGG19; mGPT; Jamba 1.5; StarCoder; MARS5; OpenFold; AceGPT; FinGPT; MAP-Neo; Saiga; Fish Speech 1.5; Kandinsky 3.0; Falcon-180B; LLaVA-1.5 and LLaVA-NeXT (LLaVA-1.6) now give sizes, training data, context lengths and results from the developers’ documentation, replacing fragments and unsourced claims such as “Top open model at release”
• Primary sources (4) — mGPT pointed at the unrelated Kandinsky 3 page and now links to its own model card; VGG16 and VGG19 now link to the Oxford project page that releases them; LLaVA 1.6 now links to the LLaVA-NeXT announcement

1,176 models · 329 vendors · 32 fields corrected on 20 models

Corrected

• 20 entries checked — Entries whose primary source was a bare GitHub or Hugging Face organization page, checked against each model’s own card, README, paper or release note; 13 descriptions rewritten and 19 primary sources replaced
• Descriptions (13) — DeepSeek-R1-0528, DeepSeek-V3-0324 (was dated March 2024; released March 2025), DeepSeek-V2.5, DeepSeek-Coder-V2, DeepSeek-V2, DeepSeek Coder, the DeepSeek chat app, DeepSeek-V4.1-Flash (now covers the October 17 V4.1-Flash-1017 update), Zhipu’s GLM-4-0414 series, Moxin-7B, InstaDeep’s Nucleotide Transformer, BAAI’s BGE and OuteTTS now give sizes, training data and capabilities from official sources, replacing fragments and unsourced claims such as “Started the Chinese AI price war”
• Primary sources (19) — Eight DeepSeek entries, GLM-4.5, GLM-4, GLM-4-0414, Depth Anything V1 and V2, Ring-1T-preview, Spectra 1.1, Moxin-7B, Nucleotide Transformer, BGE and OuteTTS now link to the model’s own card, repository, paper or homepage instead of an organization page

1,176 models · 329 vendors · 43 fields corrected on 22 models

Corrected

• 22 entries checked — Qwen and Yi entries checked against the developers’ own blog posts, model cards, READMEs and documentation; 21 descriptions rewritten and all 22 primary sources replaced
• Descriptions (21) — Alibaba’s Qwen, Qwen1.5, Qwen-72B, Qwen-Audio, Qwen-VL, Qwen-VL-Max, Qwen-VL-Plus, Qwen2, Qwen2.5, Qwen2.5-72B, Qwen2.5-Coder, Qwen2.5-Coder-32B, Qwen2-VL, GTE-large and the Qianwen assistant (formerly Tongyi Qianwen); 01.AI’s Yi-1.5, Yi-34B, Yi-6B, Yi-Coder, Yi-Large and Yi-Lightning now give sizes, training data, context lengths and licenses from official sources, replacing fragments and unsourced claims such as “Competitive with Western models” and “GPT-4 class”
• Primary sources (22) — All 22 pointed at the bare QwenLM or 01-ai GitHub organization page; each now links to the model’s own announcement, model card, repository, documentation or technical report

1,176 models · 329 vendors · 53 fields corrected on 47 models

Corrected

• 47 entries checked — More descriptions checked against each project’s paper, model card or GitHub README; all 47 rewritten
• Descriptions (47) — Meta’s ConvNeXt, MAE, DINO, DeiT, RoBERTa, ResNeXt-101, Detectron2, MMS and NLLB-200; Microsoft’s Swin Transformer; OpenAI’s CLIP and Jukebox; SciBERT, ClinicalBERT, ProGen, ProtTrans, MEDITRON, WizardLM, WizardCoder, CogVLM, CogVLM2, DeepSeek-Coder-V2-Lite, DeepSeek-VL, DeepSeek-VL2, ChatGLM-6B, YOLOv5, RFdiffusion, StableLM, DeepFloyd IF, Wonder3D, AUTOMATIC1111’s web UI, InvokeAI, MetaVoice-1B, Open-Sora, AnimateDiff, PhotoMaker, InstantID, Tortoise, Coqui TTS, XTTS v2, Donut, ColPali, GPT4All, Instant-NGP, ColBERT v2, MLX and Ferret now give sizes, training data and capabilities from the developer’s documentation, replacing fragments and unsourced claims such as “Most popular”
• Primary sources (6) — CogVLM and CogVLM2 now link to their own repositories instead of CogVideo’s; DeepSeek-Coder-V2-Lite, DeepSeek-VL, DeepSeek-VL2 and ChatGLM-6B now link to their own model cards instead of an organization page

1,176 models · 329 vendors · 47 fields corrected on 47 models

Corrected

• 47 entries checked — Descriptions of research and classic models checked against each model’s original paper; all 47 rewritten
• Descriptions (47) — Ai2’s Tülu 3 and ELMo; Microsoft’s Phi-3.5 series, Kosmos-2, VALL-E, VALL-E X, NaturalSpeech 3 and ResNets; Apple’s Apple Intelligence foundation models; Google and Google DeepMind’s Med-Gemini, Gecko, RT-2, OWLv2, OWL-ViT, SigLIP, ViT-22B, ViT-G/14, ViT-Huge/14, Flan-T5, UL2, ByT5, Switch Transformer, mT5, T5-3B, ELECTRA, ALBERT, EfficientNet, MobileNetV3, NASNet, Xception, Inception v3, GoogLeNet and the original Transformer; Meta’s Emu and CM3Leon; Baidu’s ERNIE 3.0 Titan; OpenAI’s Codex and DALL·E; plus YOLOv3, ModelScope T2V, DenseNet, Transformer-XL and RWKV-6 now say what each model is and what its paper reported, replacing fragments like “ViT-22B. Google Vision/CV model.” and unsourced claims like “Foundation of modern AI”
• GPT-3 base models (3) — ada, babbage and curie now give their former API prices and their Jan 4, 2024 shutdown, from OpenAI’s deprecations page

1,176 models · 329 vendors · 57 fields corrected on 52 models

Corrected

• 52 entries checked — Descriptions of open models and tools checked against each developer’s official model card or GitHub README; all 52 rewritten
• Descriptions (52) — Ai2’s Tülu 3 405B, TII’s Falcon 2, Stability AI, Microsoft, EleutherAI, Baichuan, Genmo, Nous Research, InternLM, PixArt, TinyLlama, Hugging Face’s Parler-TTS, Zhipu AI’s CogVideoX, Google DeepMind’s RecurrentGemma, xAI’s Grok-1, Meta (Nougat, SeamlessM4T, AudioCraft, MusicGen, Segment Anything, EnCodec, ESM-2, ESMFold, OPT-175B), OpenAI (Shap-E, Point-E), ControlNet, llama.cpp, whisper.cpp, F5-TTS, Bark, Moshi, Dia, TripoSR, Depth Anything, YOLOv10, Grounding DINO, Stanford Alpaca and others now give sizes, training data and capabilities from the developer’s own documentation, replacing fragments and unsourced claims such as “Largest open aligned” and “First strong open GPT alternative”; Falcon 2 no longer says Apache 2.0 or vision-language
• Primary sources (4) — CroissantLLM, Moxin-7B and OpenHermes 2.5 now link to the model’s own page instead of an organization or profile page; PixArt-Σ now links to its own repository instead of PixArt-α’s
• Licenses (1) — Baichuan 2 now states its commercial-license conditions: an email application, under 1 million daily active users, and no software or cloud providers

1,176 models · 329 vendors · 46 fields corrected on 45 models

Corrected

• 45 entries checked — More model descriptions checked against each developer’s official model card; all 45 rewritten
• Descriptions (45) — Open models from OpenAI (Whisper large-v3 and v3-turbo), Mistral AI, Apple, Stability AI, Microsoft, Google, AI21, BAAI, BigCode, Hugging Face, Salesforce, Tencent, NVIDIA, Snowflake, Nous Research, TII, Together, Replit, LMSYS, LLM360, BigScience and others now give sizes, context lengths, training data and capabilities from the developer’s card, replacing fragments and unsourced claims such as “8x speedup”, “90% ChatGPT quality” and “Top retrieval embedding”
• Licenses (1) — Stable Zero123 now distinguishes the non-commercial original from the commercially licensed Stable Zero123C

1,176 models · 329 vendors · 62 fields corrected on 39 models

Corrected

• 39 entries checked — Model descriptions checked against each developer’s official model card; all 39 rewritten, and 22 now cite the model’s own card as their primary source
• Descriptions (39) — Qwen3 (8), Phi-3 and Phi-3.5 (7), Llama 3, 3.1, 3.2 and 2 (14), Code Llama (5), EXAONE (3) and two DeepSeek-R1 distilled models now give parameters, context length, training data and capabilities from the developer’s card, replacing fragments and unsourced claims such as “Near GPT-4 class” and “The largest open-source model on Earth”
• Sources (22) — Primary sources that pointed to Meta’s general Llama page or a GitHub organization now point to the model’s own card
• Licenses (1) — DeepSeek-R1-Distill-Llama-8B now notes it derives from Llama 3.1 8B, originally under the Llama 3.1 license

1,176 models · 329 vendors · 27 fields corrected on 12 models

Corrected

• 12 entries checked — Apps that had no price or only a placeholder were checked against each vendor’s own pricing page or help center; all 12 now show a sourced price or how the product is sold
• Prices (12) — Reve, Boomy, Craiyon, Kaiber, Vidu, Replika and Moises now show their plans; Resemble AI no longer sells voice products and now prices deepfake detection; Upstage Solar Pro shows Solar Pro 4’s API price and the Oct 30, 2026 end of service for Solar Pro 3 and Pro 2; CrowdStrike Charlotte AI gives qualifying customers 50 free AI credits a month; Harvey and DataRobot publish no prices
• Free tier (1) — Kaiber has no free plan, only a $5 five-day trial
• Descriptions (2) — Resemble AI is now described as a deepfake-detection company; Upstage Solar Pro now describes the current Solar Pro 4

1,176 models · 329 vendors · 117 fields corrected on 81 models

Corrected

• 81 entries checked — Open-model licenses checked against each model’s Hugging Face repo or GitHub LICENSE; 35 entries corrected and 46 now cite the license page their price comes from
• Licenses (32) — 27 Llama, Llama Guard and Code Llama entries now name their exact license (Llama 4, 3.3, 3.2 or 3.1 Community License, Meta Llama 3 Community License or Llama 2 Community License) instead of a generic “Llama license”, and six drop “also free in the Meta AI app” because Muse Spark now powers Meta AI; the original LLaMA was a noncommercial research release, not free open weights; Chameleon, Fuyu-8B, DeepFloyd IF and Nemotron 3 Nano name their exact licenses
• Availability (3) — Ring-1T-preview and ICONN-1 are no longer public on Hugging Face; Meta Muse Spark now carries its Meta Model API price and free Meta AI app use
• Sources (46) — 46 other open-model entries, from Gemma 4 and Qwen-Image to Aider and YOLOv10, were checked against their licenses and now cite them

1,176 models · 329 vendors · 20 fields corrected on 15 models

Corrected

• 15 entries corrected — Descriptions that still called superseded models their vendor’s flagship, and plan prices that had changed, checked against each vendor’s model list, docs or pricing page
• Descriptions (11) — Nano Banana 2, Grok 4.20, GLM-5, Gemma 3 27B, Pixtral Large, Llama 3.2 90B Vision, FLUX.1 Pro and Ideogram 2.0 no longer read as their vendor’s current flagship (Google now recommends Nano Banana 2.1; xAI calls Grok 4.7 its flagship; Z.ai lists GLM-5.3; Gemma 4, Llama 4, FLUX 3 and Ideogram 4.5 have followed); Kimi K2.6 notes Kimi K3; Eleven v3 is now ElevenLabs’ previous generation after Eleven v4; ElevenLabs Turbo v2 was English-only, not 29 languages, and is deprecated
• Prices (8) — Ideogram 1.0, 2.0, 3.0 and 4.0 drop a Basic plan Ideogram no longer sells (Plus $20, Pro $60, Team $30 per user); ElevenLabs, Eleven v3 and Turbo v2 now show Starter at $6, not $5, and the full plan range; FLUX.1 Pro is $0.05 per image

1,176 models · 329 vendors · 60 fields corrected on 46 models

Corrected

• 46 entries checked — Prices for entries that cited no pricing page checked against each vendor’s own pricing page, model page or price list; 13 were wrong or out of date, and 33 now cite the page their price comes from
• Corrected (13) — Muse Spark 1.3, 1.2 and 1.1 now carry Meta’s Standard-mode API price ($1.25/$4.25 per 1M tokens) and Contributor-mode rates; Muse Image is now on the Meta Model API at $0.01 per image (was listed as having no public API); LongCat-2.0 costs $0.30/$1.20 per 1M tokens, not ~$0.038; Bland Speech v3 is $0.04 per 1,000 characters with 50,000 free (was $0.015, no free tier); FLUX 3 adds its per-second video and per-image prices; GPT-Transcribe and GPT-Live-Transcribe now give OpenAI’s per-minute prices; Julius no longer has a Max plan; Pika 1.0 and Pika 2.0 are off Pika’s model list and its old Basic/Standard plans are gone; Windsurf SWE-1 is off the Devin Desktop model list
• Sources (33) — 33 other prices, from Base44 and Brave to Warp and You.com, were confirmed unchanged and now cite their pricing pages

1,176 models · 329 vendors · 32 fields corrected on 19 models

Corrected

• 19 entries corrected — Gemini, Nano Banana, Imagen, Veo and Lyria entries checked against Google’s Gemini API pricing and deprecations pages and Google Cloud’s price list and model lifecycle page; one price was off by a factor of 1,000, and two Flash models now redirect to newer ones on the Gemini API
• Prices (9) — Gemini 3.6 Flash is $0.75/$3.75 per 1M tokens through Dec 31, 2026 (was listed at $1.50/$7.50); NanoBanana 2 Lite costs $0.0336 per 1K-resolution image, not per 1,000 images; Gemini 3.8 Flash Cyber now carries its published Google Cloud price ($1.50/$7.50); Gemini 3.1 Pro adds its $2/$12 API price; Nano Banana and OmniFlash now list their per-token API prices; Gemini 3.1 Flash-Lite, 2.5 Flash and 2.5 Flash-Lite add audio-input prices
• Availability (7) — Gemini API requests for Gemini 3.5 Flash and 3.7 Flash now route to 3.6 Flash and 3.8 Flash (both still sold on Google Cloud); Gemini 3 Pro’s API preview shut down March 9, 2026; Imagen 4 and Imagen 3 are shut down in the Gemini API; Veo 3.1’s API previews may shut down from Oct 22, 2026; Gemini 2.5 Pro, Flash and Flash-Lite are now open only to developers who used them before
• Sources (3) — Gemini 3.8 Flash, Gemini 3.5 Flash-Lite and Lyria 3.5 now cite Google’s price list

1,176 models · 329 vendors · 16 fields corrected on 10 models

Corrected

• 10 entries corrected — Claude and OpenAI API prices checked against Anthropic’s pricing page and OpenAI’s model pages; two prices had changed and three Claude models had published rates we listed as unpriced
• Anthropic (5) — Claude Sonnet 5 is still $2/$10 per 1M tokens, not the $3/$15 the old entry said would apply after Aug 31, 2026; Claude Mythos 5.1 and Claude Mythos 5 now carry their published $10/$50 rates (access restrictions unchanged); Claude Fable 5 adds its $50 output price; Claude Opus 4.6 adds its $5/$25 API price
• OpenAI (5) — GPT-5.6 Sol is now $4/$20 (promotional pricing through at least Nov 21, 2026; was $5/$30) and GPT-5.6 Terra $2/$12 (was $2.50/$15); GPT Image 2.5 Flare and Sunburst add their $5 text-input price; GPT-6 Astra now cites its model page

1,176 models · 329 vendors · 12 fields corrected on 9 models

Corrected

• 9 entries corrected — Descriptions with stale “flagship” and “recommended” claims are updated, and xAI prices now include cached-input and long-context rates from xAI’s own price list
• Descriptions (5) — GPT-5 said details were still emerging; it now describes the shipped model and says it is deprecated in favor of GPT-6 Astra. GPT-5.5 is no longer called OpenAI’s current flagship; Grok 4.3 no longer says xAI recommends it (Grok 4.7 now is); Qwen3-Max is no longer called Alibaba’s largest model (Qwen3.8-Max has 2.4 trillion parameters); Wan 3.0 drops speculation about open weights
• xAI prices (5) — Grok 4.7, 4.6, 4.5, 4.3 and 4.20 now include cached-input prices and the higher rates xAI charges for prompts of 200K tokens or more

1,176 models · 329 vendors · 11 fields corrected on 11 models

Corrected

• 11 descriptions corrected — Descriptions that still called a superseded model the latest, most capable or best now say what replaced it, from each vendor’s own model list or version page
• Claude (3) — Claude Opus 4.7, Claude Opus 4.6 and Claude Haiku 4.5 are superseded by Anthropic’s current Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5
• Midjourney (3) — V7, V6.1 and V6 now give their release and default-version dates from Midjourney’s Version page; V8.2 has been the default since July 24, 2026
• Others (5) — Llama Guard 3 (succeeded by Llama Guard 4), ERNIE 4.0 (succeeded by the open ERNIE 4.5 family), Ideogram 3.0 (succeeded by Ideogram 4.0), GLM-4 (succeeded by GLM-4.6, GLM-4.7 and the GLM-5 series) and Lyria 3 (succeeded by Lyria 3.5)

1,176 models · 329 vendors · 12 fields corrected on 11 models

Corrected

• 11 descriptions corrected — Generic “model/tool for … workflows” templates and internal editing notes are replaced with descriptions from each vendor’s own launch post, model page or model card
• Wrong or generic (6) — Gemini Diffusion was described as an image generator but is Google DeepMind’s experimental text diffusion language model (about 1,479 tokens/sec; demo only), and its primary source now points to its own page; Cosmos 1.0, Apple DiffuCoder, DeepCoder-14B-Preview, Windsurf SWE-1 and IBM Granite Code Models get real descriptions
• Internal notes and stale claims (5) — DeepSeek-V3.2 no longer says the legacy deepseek-chat and deepseek-reasoner API names route to it (they pointed to V4-Flash and were discontinued on July 24, 2026); Qwen3.5 gets a real description (Qwen3.5-397B-A17B, 201 languages); Kling 3.0, Runway Gen-4.5 and Luma Ray3 lose an internal “Backfill row” editing note

1,176 models · 329 vendors · 24 fields corrected on 16 models

Corrected

• 16 entries corrected — Open-weight models now say their weights are free and name the license, and a hosted-only model is no longer listed as open
• Prices (5) — Whisper: MIT weights, OpenAI API $0.006/min (deprecated); Inkling: Apache 2.0 weights, Tinker API $1.00/M input, $4.05/M output; Inkling-Small: Apache 2.0 weights, Tinker API $0.30/$1.20; Hy-MT2-30B-A3B and Hy-MT2-1.8B: Apache 2.0 weights, API through OpenRouter, served by Tencent Cloud
• Qwen3.6-Plus — Listed as open weights, but Qwen’s launch post calls it the hosted model on Alibaba Cloud Model Studio, and only the smaller Qwen3.6 models were open-sourced: now closed and cloud-only, with a real description
• Free tier (11) — Whisper, Granite 4.0 Tiny Preview, Granite 3.3, 3.2 and 3.1 8B, Granite Code, Bamba-9B-v2, Llama-3.1-Nemotron-Ultra-253B, Llama-3.3-Nemotron-Super-49B, Sesame CSM and Falcon were marked as having no free tier, though their weights are a free download

1,176 models · 329 vendors · 28 fields corrected on 11 models

Corrected

• 11 entries updated — Blank or vague prices now give real figures from each vendor’s own pricing page, help center or model documentation, and five entries with wrong details are corrected
• Prices (6) — Poe: free tier, subscriptions from $4.99/mo, add-on points from $30 per 1M; Snapchat My AI: free to all Snapchat users (Snapchat+ adds extra options); Yi-Large: $3/M input and output on 01.AI’s API; Yi-Lightning: CNY 0.99/M tokens on 01.AI’s China API, where the endpoint now also routes to DeepSeek-V3 and Qwen3-30B-A3B; Kling 2.5 Turbo: $0.042/sec at 720p and $0.07/sec at 1080p on Kling’s API; Fugaku-LLM: open weights under the Fugaku-LLM Terms of Use, with a new description
• Granite 13B — Listed as open weights, but IBM never published them: now closed and watsonx.ai-only, where granite-13b-chat-v2 runs on dedicated deploy-on-demand hosting and granite-13b-instruct-v2 was withdrawn on Oct 15, 2025; new description
• Other corrections (4) — Amazon Titan Text is no longer on Amazon Bedrock’s supported model list; Med-Gemini is now marked as an unreleased research model; Relevance AI no longer offers a free plan; Cosmos Nemotron 34B is a vision language model, not a video generator. Snapchat My AI’s description now names Google’s Gemini models alongside OpenAI’s GPT

1,176 models · 329 vendors · 25 fields corrected on 14 models

Corrected

• 14 entries corrected — Vague values such as “See official pricing page”, “usage-based pricing” and “Licensing/API varies” now give real prices or say the model is no longer sold, from each vendor’s own pricing page
• Plan prices (3) — Wordtune: Free; Advanced $13.99, Unlimited $19.99/mo ($6.99 and $9.99 billed annually), replacing a value that described AI21’s developer API; Fliki: Free; Standard $28, Premium $88/mo per user ($21 and $66 billed annually); Midjourney V8.1: Basic $10 to Mega $120/mo, with HD images using 1.3 GPU-minutes vs 0.8 for SD
• No longer sold (5) — Stable Audio 2.0 is off Stability AI’s API price list (Stable Audio 2.5 and 3.0 cost $0.08 per generation); Inception’s original Mercury and Mercury Coder Small are off its model list (now Mercury 2.5, Voice, Decide, 2 and Edit 2); Jurassic-1 Jumbo and Jurassic-2 are off AI21’s price list, which now sells only Jamba
• Zhipu GLM (6) — GLM-5.3 (GLM-5.3 license) and GLM-5.3-Flash (MIT) now name their open-weight licenses and cached prices ($1.40/$4.40 and $0.15/$0.50 per 1M tokens); GLM-4.5 ($0.60/$2.20) and both GLM-4-0414 entries ($0.10/$0.10) add their Z.ai API prices; GLM-5.2 now cites Z.ai’s price list

1,176 models · 329 vendors · 18 fields corrected on 9 models

Corrected

• 9 entries corrected — Older open-weight Mistral models said only “Free (open weights, Apache 2.0)”; each now also says its hosted Mistral API version is retired, with the date from Mistral’s own model page
• Retired from the Mistral API — Mixtral 8x22B, Mixtral 8x7B, Mixtral 8x7B Instruct, Mistral 7B, Mistral-7B-v0.2 and Mistral-7B-Instruct-v0.3 (March 30, 2025); Codestral Mamba (June 6, 2025); Pixtral 12B (Dec 31, 2025); Mathstral (retired, no date given). The weights remain Apache 2.0

1,176 models · 329 vendors · 28 fields corrected on 18 models

Corrected

• 18 entries corrected — Placeholders such as “Kimi subscription (China-focused)”, “YandexGPT subscription (Russia-focused)”, “N/A (enterprise/regional)” and “Alibaba Cloud Model Studio” now give a real price or say the model is no longer sold, from each vendor’s own price list or model list
• Kimi (4) — Kimi Chat: Free; Plus $19, Pro $39, Max $99, Ultra $199/mo ($15 to $159/mo billed annually); Kimi K3: open weights (Kimi K3 license), API $3/M input, $15/M output, and the claim that its license fixes prices across providers is removed; Kimi k1.5 is not on Kimi’s API price list; Kimi K2.7 Code now cites the API price page
• Yandex (4) — YandexGPT, YandexGPT 2, YandexGPT 3 Pro and YandexGPT 3 Lite are retired: Yandex AI Studio now offers only YandexGPT Pro 5 and 5.1 and YandexGPT Lite 5
• Tencent (3) — Hunyuan T1 and Hunyuan Turbo S are no longer on Tencent Cloud’s Hunyuan price list, as Hunyuan moves to TokenHub (Hy3 and Hy4 Preview); Hunyuan-A13B adds its Tencent Cloud API price (CNY 0.5/M input, CNY 2/M output)
• Alibaba (7) — Qwen2.5-Max is retired (Alibaba Cloud says Qwen2.5-series models are no longer available); Qwen3-Max, Qwen3-Max-Preview, QwQ-Max-Preview, Qwen-VL-Max and Qwen-VL-Plus are no longer on Alibaba Cloud’s Model Studio price list; Qwen3.8-Max-0902 now cites the price list

1,176 models · 329 vendors · 49 fields corrected on 24 models

Corrected

• 24 entries corrected — Fifteen models with no price now carry one from the vendor’s own pricing page or docs, and nine more get corrected prices, licenses, retirement dates or sources
• Missing prices filled (15) — Gamma (Free; Plus $12, Pro $25, Ultra $100/mo); NightCafe ($12.49 to $62.49/mo; 5 free credits a day); Glif ($10, $30, $100/mo); Reve 2.1 (Free; Lite $7.99, Pro $19.99, Teams $59/user/mo); Magnific AI (in Magnific, formerly Freepik, plans from $20/mo); PixVerse (Free; $10, $30, $60/mo); Kling 3.0 (API from $0.084/sec at 720p); Recraft V3 (API $0.04 per raster, $0.08 per vector image); Reflect (free on Mac); GeoGPT (open weights, GeoGPT license); MAI-Transcribe-1.5 ($0.36 per audio hour); MAI-Voice-2 ($22 per 1M characters); MAI-Code-1-Flash (retired from GitHub Copilot on Sept 10, 2026); Med-PaLM 2 (retired with Google Cloud’s MedLM on Sept 29, 2025); Robostral Navigate (not publicly priced)
• Outdated or wrong labels fixed (2) — Robostral Navigate was listed as open weights with an API, but Mistral has published neither and offers it through its team, so it is now closed and not publicly available; Reflect now ships as Reflect Open, an open-source (MIT) app with local Markdown files, so it is now open source and local
• Mistral corrections (4) — Mistral Large (February 2024) was retired from the API on June 16, 2025 and had a 32K, not 128K, context window; Codestral (May 2024, 22B) is non-production open weights (MNPL) and its API version was retired on June 16, 2025; Mistral NeMo’s API was retired on July 31, 2026; Le Chat lists Free, Pro $14.99/mo and Team $24.99/user/mo
• Also fixed — MAI-DS-R1 now shows its Microsoft Foundry price ($1.35/M input, $5.40/M output); MAI-Image-2.5 Pro is restated in the standard format; MAI-Image-2.5, MAI-Thinking-1 and MAI-Cyber-1-Flash now cite Azure’s price page

1,176 models · 329 vendors · 24 fields corrected on 11 models

Corrected

• 11 entries corrected — Prices and licenses for open models now come from the vendor’s own price list and official repository, and three entries that showed the wrong product’s price or license status are fixed
• Open-model prices and licenses (8) — GLM-5.1 ($1.40/M input, $4.40/M output), GLM-5 ($1.00/$3.20) and GLM-4.7 ($0.60/$2.20) are MIT open weights on the Z.ai API; MiniMax-M2.1 (modified MIT; $0.30/$1.20); FLUX.1 Kontext ($0.04 [pro], $0.08 [max] per image); FLUX.2 (from $0.03 [pro], $0.05 [flex], $0.07 [max] per image); both FLUX [dev] models are non-commercial open weights; Gemini CLI (Apache 2.0; 1,000 free requests a day with a Google account, not 1M tokens); LangFlow (MIT)
• Wrong entries fixed (3) — Bark showed Suno’s music-app prices but is Suno’s free MIT-licensed text-to-audio model; DeepSeek-GRM-27B showed a DeepSeek API price but is free open weights under the Gemma license with no hosted API; codex-1 was marked open source, but OpenAI describes it as a version of o3 that powers the cloud Codex agent, so it is now proprietary and cloud-only

1,176 models · 329 vendors · 24 fields corrected on 12 models

Corrected

• 12 entries now give real prices or licenses — These said only “API”, “subscription tiers”, “see Google pricing” or “Paid API”; each now carries the figures from the vendor’s own price list, plan page or repository
• Google API prices (4) — Gemini 3 Flash ($0.50/M input, $3/M output, with a free tier); Veo 3.1 ($0.40 per second at 720p or 1080p, $0.60 at 4K; Fast from $0.10, Lite from $0.05); Nano Banana 2 ($0.067 per 1K image); Nano Banana Pro ($0.134 per 1K or 2K image, $0.24 at 4K)
• Plan prices (5) — Suno (Pro $10/mo, Premier $30/mo); Eleven Music via ElevenLabs plans (Starter $6/mo to Business $990/mo, about 900 credits per minute of music); Recraft V4 (Basic $12, Pro $20, Teams $22 per seat); Midjourney V1 Video (every plan, $10 to $120/mo); Character.AI (c.ai+ $9.99/mo or $94.99/yr)
• Open-weight licenses (3) — SAM 3 is open weights under the SAM License, not a paid API; Segment Anything and Moondream 2 are Apache 2.0
• Also fixed — SAM 3’s placeholder description now describes the model, including the SAM 3.1 checkpoints released in March 2026

1,176 models · 329 vendors · 38 fields corrected on 22 models

Corrected

• 11 open models no longer say N/A for price — Their price said only “N/A (research)”, “N/A (internal research)” or “N/A (enterprise/research)”, but each has free downloadable weights; each now names the license on its official Hugging Face or GitHub page
• Non-commercial (4) — EXAONE Deep and EXAONE-3.5 (EXAONE AI Model License 1.1 - NC), EXAONE 4.0 (1.2 - NC) and ResNeXt-101 (pretrained weights CC BY-NC 4.0, code BSD)
• Apache 2.0 and MIT (7) — Both Sailor2 entries, DINOv2 and ViT-Huge/14 (Apache 2.0); DeepSeek-TNG-R1T2-Chimera, EVA-01 and wav2vec 2.0 Large (MIT)
• Also fixed — Free Tier is now Yes for the five of these that said N/A; 11 Google research models no longer cite the Gemini API price list as their pricing source, since it does not mention them

1,176 models · 329 vendors · 32 fields corrected on 12 models

Corrected

• 12 entries now give a real price or their current status — These said only “API pricing”, “Usage-based API pricing”, “Subscription” or “Licensing/API varies”; each now carries the figure or status from the vendor’s own pricing page, docs, changelog or Hugging Face repo
• Prices named (5) — FLUX 1.1 Pro ($0.04 per image) and FLUX 1.1 Ultra ($0.06 per image) on the BFL API; both voyage-3 entries ($0.06 per million tokens; Voyage now lists voyage-3 among older models with no free tokens, so Free Tier is now No); Udio (free plan, Standard $10/mo, Pro $30/mo)
• Open weights (2) — Jina Reranker v2 is open weights under CC BY-NC 4.0 (non-commercial), with a paid Jina API, and is now marked open source; Featherless Qwerky-72B, now QRWKV-72B, is free open weights under the Qwen license, with Featherless hosting from $25/mo
• Retired or replaced (5) — Stable Diffusion 1.5, 2.0 and 2.1 are now Discontinued, since their official Hugging Face repos are gone and Stability’s API does not offer them; DreamStudio now redirects to Stability AI’s Brand Studio; Hume EVI 2 was deprecated on August 30, 2025, and Hume is shutting down its whole EVI and TTS APIs on November 13, 2026
• Also fixed — FLUX 1.1 Pro is no longer marked partly open source (FLUX.1 [schnell] is a separate model); one voyage-3 description no longer calls it Voyage’s latest model; broken BFL and Voyage pricing links were replaced

1,176 models · 329 vendors · 42 fields corrected on 19 models

Corrected

• 19 more entries now name their license or status — These listed a paid API or platform pricing for free-to-download models, or a vague “Free” or “No paid pricing” label; each now names the license on its official Hugging Face or GitHub page, or its correct status
• Open models listed as paid APIs (6) — Cerebras-GPT was released as open models, not as the Cerebras Inference API, and its official weights are no longer published; Baichuan 2 is open weights under the Baichuan 2 Community License (free commercial license on request); RedPajama is an open dataset whose subsets keep their source licenses; RedPajama-INCITE-7B, Snowflake Arctic and Snowflake Arctic Embed are Apache 2.0
• Licenses named (10) — Ling-3.0-Flash-Fin, GPT4All and V-JEPA 2 (MIT); Muse Glimmer, Neural Chat and Neural Chat 7B (Apache 2.0); SOLAR 10.7B (Apache 2.0 base, CC BY-NC 4.0 Instruct); StyleTTS 2 (MIT code, and its pre-trained models require telling listeners the speech is synthesized); YOLOv5 and YOLOv8 (AGPL-3.0, with an Ultralytics Enterprise License for closed-source commercial use)
• Non-commercial terms (3) — EXAONE 4.5 (EXAONE AI Model License 1.2 - NC), AudioLDM 2 (CC BY-NC-SA 4.0) and Jina Embeddings v3 (CC BY-NC 4.0; Jina also sells API access)
• Also fixed — Cerebras-GPT is now marked open source with local deployment; RedPajama’s deployment type is now Local (open dataset); YOLOv8 is no longer called the latest Ultralytics YOLO, since Ultralytics now names YOLO26

1,176 models · 329 vendors · 153 fields corrected on 76 models

Corrected

• 76 more entries now name their license or status — These said only “Free (open-source)”, “Free (open weights/code)”, “Free (research)” or “No paid pricing”; each now names the license on its official Hugging Face or GitHub page, or its correct status
• Apache 2.0 (37) — OLMo, OLMo 2 32B, FlexOlmo, Pythia, GPT-J, GPT-NeoX, Falcon, TinyLlama, ModernBERT, SAM 2, Wan, AnimateDiff, Grounding DINO, WildGuard, SciBERT, ELMo, Comma v0.1-2T, METAGENE-1, Sesame CSM, RWKV-6, RWKV-7 Goose, INTELLECT-1, INTELLECT-2, Hermes-2-Pro-Mistral-7B, OpenHermes 2.5, Dolphin 2.5 Mixtral, the six K2 Horizon models, K2-Think, Evo 2, s1-32B, Sky-T1-32B-Preview and Bamba-9B
• MIT (12) — BGE, BGE-Large, BGE-M3, EVA-02, BLIP-2, GPT-Neo, mmBERT, ClinicalBERT, ColBERT v2, YuLan-Mini, Nous Capybara and TNG DeepSeek Chimera
• Llama licenses (6) — Tulu 3 and both Tulu-3-405B entries (Llama 3.1), ELYZA-japanese-Llama-2-7b (Llama 2), and Hermes 4 and DeepHermes 3 Preview, whose Llama-based sizes use the Llama 3 license and other sizes Apache 2.0
• Other licenses (13) — VGG16 and VGG19 (CC BY 4.0), Helium-1 (CC BY-SA 4.0), Moshi (weights CC BY 4.0), SAM 3D (SAM License), Stable Diffusion 3.5 (Stability AI Community License), Stable Diffusion XL (CreativeML Open RAIL++-M), StarCoder 2 (BigCode OpenRAIL-M), Falcon 2 (TII Falcon License 2.0), ControlNet (OpenRAIL weights, Apache 2.0 code), DenseNet (BSD 3-Clause), YOLOv3 (public-domain code) and GloVe (public-domain vectors)
• Open-source software and non-commercial terms (5) — llama.cpp (MIT), OpenAssistant (Apache 2.0) and Coqui TTS (MPL 2.0; its XTTS-v2 weights are non-commercial only); Fish Speech 1.5 (CC BY-NC-SA 4.0) and Alpaca (research only) are free only for non-commercial use
• Status fixes (3) — Ollama now lists cloud plans (Pro $20/mo or $200/yr, Max $100/mo, Team $500/mo in early access) on top of free local use; Semantic Scholar is a free service, not open-source software; Nous DisTrO 15B’s weights were never published

1,176 models · 329 vendors · 66 fields corrected on 26 models

Corrected

• 17 models now name their license — Apache 2.0: EfficientNet, EfficientNetV2-XL, MobileNetV3, NASNet, Inception v3, GoogLeNet/Inception v1, Word2Vec and Mistral-7B-v0.2. MIT: Xception, ResNet-50, ResNet-101, ResNet-152, BEiT and EnCodec. Falcon-180B: Falcon-180B TII License. Both Transformer entries: open-source code under Apache 2.0
• Retired by Databricks (4) — DBRX, DBRX Instruct, MPT-7B and MPT-30B: Databricks retired them from its model-serving platform and no longer publishes the weights
• Weights no longer published (2) — Dolly 2.0 (Databricks removed it from Hugging Face) and WizardLM 2 (Microsoft pulled the weights soon after its April 2024 release)
• Not released (3) — Orca and ViT-22B weights were never published, and DolphinGemma is not yet released (Google says it will be openly available on release); their open-source, deployment and free-tier fields now say so
• Descriptions (4) — DolphinGemma now describes the dolphin-communication model; Mistral-7B-v0.2 no longer claims sliding-window attention, which v0.2 dropped; MPT-7B’s 65K context is credited to its StoryWriter variant; the Transformer entry loses a stray quotation mark

1,176 models · 329 vendors · 102 fields corrected on 51 models

Corrected

• 51 research models now name their license — These models said only “Free (open weights)”; each now names the license on its official GitHub or Hugging Face repository
• Apache 2.0 (22) — SAM, DINO, DeiT, Detectron2, YaLM 100B, BERT, ALBERT, ELECTRA, T5, FLAN-T5, mT5, ByT5, UL2, Switch Transformer, ViT, SigLIP, OWL-ViT, OWLv2 and XLNet
• MIT (14) — CLIP, GPT-1, Shap-E, Point-E, ConvNeXt, Swin Transformer, ESM-2, ESMFold, RoBERTa, E5, BioGPT, DeBERTa v3 and PubMedBERT
• Non-commercial terms (12) — Seamless M4T, MAE, NLLB-200 and Galactica (CC BY-NC 4.0), AudioCraft and MusicGen (weights CC BY-NC 4.0, code MIT), Nougat (weights CC-BY-NC), ImageBind and LayoutLM v3 (CC BY-NC-SA 4.0), Jukebox (OpenAI Noncommercial Use License), and OPT and OPT-175B (OPT-175B License, non-commercial research only; OPT-175B weights on request) now say they are free only for non-commercial use
• Other — AlphaFold 2 parameters are CC BY 4.0 (code Apache 2.0), GPT-2 is under a modified MIT license, and Whisper.cpp is MIT-licensed open-source software

1,176 models · 329 vendors · 220 fields corrected on 110 models

Corrected

• 110 models now name their license — These models said only “Free (open weights)”; each now names the license on its official Hugging Face or GitHub repository
• Apache 2.0 (34) — gpt-oss, Qwen3-Coder, Mistral 7B, Mixtral 8x7B and 8x22B, Mistral NeMo, Mathstral, Codestral Mamba, Pixtral 12B, Grok-1, the Yi models, FLUX.1 Schnell, StepFun Step 3, Reka Flash 3 and 3.1, Intern-S1, InternLM3 and others
• MIT (31) — the Phi-1 to Phi-4 families, Florence-2, MAI-DS-R1, BitNet b1.58, Whisper Large v3 and v3 Turbo, CLIP, dots.llm1, MiMo, DeepCoder and others
• Gemma Terms of Use (23) — Gemma 1, 2, 3 and 3n, CodeGemma, PaliGemma 2, ShieldGemma, RecurrentGemma, T5Gemma, VaultGemma, EmbeddingGemma and TranslateGemma; MedGemma and TxGemma are under Google’s Health AI Developer Foundations terms
• Other licenses — Llama 3 and Llama 4 Scout (Meta Llama Community Licenses), Nemotron 3 Super (NVIDIA Nemotron Open Model License), Falcon 3, Falcon-H1, Falcon-Edge and Falcon2-11B-VLM (TII Falcon licenses), Stable Video Diffusion (Stability AI Community License), MiniMax-Text-01, Baichuan-M1-14B, KAT-V1-40B, TildeOpen-30b (CC BY 4.0), GPT-2 (Modified MIT) and InternLM 2.5. Qwen2.5-Coder, InternVL 2.5 and Replit Code note the models whose license differs
• Non-commercial terms — Orca 2 (Microsoft Research License, non-commercial research only), MMS (CC BY-NC 4.0) and FLUX.1 [dev] (FLUX.1 [dev] Non-Commercial License) now say they are free only for non-commercial use

1,176 models · 329 vendors · 29 fields corrected on 14 models

Corrected

• Perplexity — Perplexity Pro now lists Perplexity’s current plans: Pro $17 and Max $167 a month, and Enterprise Pro $34 and Enterprise Max $271 per seat a month, all billed annually (was “$20/mo Pro”). Sonar, Sonar Large and R1 1776 had shown that consumer price: Sonar now shows its Agent API price ($0.25/$2.50 per 1M input/output tokens plus tool fees), and Sonar Large (removed from Perplexity’s API on February 22, 2025) and R1 1776 (removed August 1, 2025) are retired
• DeepSeek — DeepSeek-V3.2, V3.2-Exp, V3.2-Speciale and R1 are MIT open weights that DeepSeek’s API no longer serves (it now runs V4.1-Flash and V4-Pro); V3.2-Speciale’s temporary endpoint ended December 15, 2025, and V3.2-Exp’s deployment no longer names DeepSeek’s app and API. DeepSeek Coder V2 is under the DeepSeek License
• Open-weight licenses — Kimi K2 (Modified MIT), MiniMax-M1 (Apache 2.0), Moonlight (MIT), Hunyuan-A13B (Tencent Hunyuan Community License, not licensed in the EU, UK or South Korea) and Dream 7B (Apache 2.0) now name their licenses instead of “Free (open weights)”

1,176 models · 329 vendors · 63 fields corrected on 36 models

Corrected

• DeepSeek — DeepSeek V3, V3-0324, R1-0528 and V3.1-Terminus showed an old $0.27/$1.10 API price, but DeepSeek’s API now serves only V4.1-Flash and V4-Pro; they now show their open-weight licenses (MIT, or the DeepSeek Model License for V3). V4-Pro shows its current price ($1.32/$3.96 per 1M tokens at peak, half off-peak); V4-Flash and V4-Flash-Vision-Exp are retired API names now served by V4.1-Flash; V2.5 is under the DeepSeek License
• MiniMax and Kimi — MiniMax M2 and M2.5 (Modified MIT weights) and M2.7 and M3 (non-commercial weights) now show their API prices ($0.30/$1.20 per 1M tokens) instead of “Free (limits); API”. Kimi K2.5 and K2 Thinking are Modified MIT open weights no longer on Kimi’s API price list; K2.6 shows its current price ($0.95/$4.00)
• Midjourney — Six Midjourney rows now list all four plans: Basic $10, Standard $30, Pro $60 and Mega $120 a month, 20% off yearly
• Open weights — HunyuanVideo (not licensed in the EU, UK or South Korea) and Hunyuan-DiT are under the Tencent Hunyuan Community License, YOLO-World is GPL-3.0 and PixArt-α and PixArt-Σ are CreativeML OpenRAIL++-M; they had said “N/A (enterprise)”
• Blank prices filled — Seedance 2.5 and Seedream 5.0 Pro (Volcengine Ark), Captions (Max $24.99, Frontier from $69.99), Meshy (Pro $20 to Ultra $100), GroqCloud (e.g. GPT-OSS-120B $0.15/$0.60), Arc Search (free), Siri, LinkedIn EON-8B (internal, not sold) and GigaChat (free API allowance for individuals)

1,176 models · 329 vendors · 96 fields corrected on 48 models

Corrected

• Licenses named — 48 open models and tools said only “Free/open weights; hosting costs vary”. They now name their license from the official model card or repository: Apache 2.0 (Apertus, Moxin, Salamandra, AceGPT, MAP-Neo, Dia, Depth Anything, Open-Sora, PhotoMaker, Tortoise TTS, InvokeAI, Jan), MIT (CroissantLLM, TripoSR, FinGPT, ProtTrans), BigCode OpenRAIL-M (StarCoder, StarCoder2), Llama licenses (Vicuna, Meditron, LLaVA 1.5, Deepthought-8B, Saiga), GPL or AGPL (Automatic1111, ComfyUI, Fooocus, Piper) and BSD (RFdiffusion)
• Non-commercial and research-only — Isaac 0.1, Nucleotide Transformer, OuteTTS, XTTS v2, the F5-TTS weights, Depth Anything V2 Base/Large/Giant and the InstantID checkpoints are non-commercial or research-only; Instella-3B is under AMD’s research-only RAIL license
• Hosted plans — ComfyUI, Flowise and Dify also show their hosted versions: Comfy Cloud from $20 a month; Flowise Cloud Free, Starter $35 and Pro $65 a month; Dify Cloud Sandbox (free), Professional $590 and Team $1,590 per workspace a year

1,176 models · 329 vendors · 31 fields corrected on 16 models

Corrected

• Runway — Gen-1 and Gen-2 were deprecated in 2025, and Runway retired Gen-3 Alpha on July 8, 2026 and Gen-3 Alpha Turbo on July 30, 2026, so all four now say “Retired (no longer sold)”. Runway (platform) and Gen-4 now show Runway’s current plans: Free with 125 one-time credits, Standard $15, Pro $35 and Max $95 a month. The original August 2022 Stable Diffusion release is free under the CreativeML OpenRAIL-M license
• ByteDance and Baidu — UI-TARS-72B and ERNIE-4.5-21B-A3B-Thinking are free open weights under Apache 2.0. Seedance now shows its Volcengine price (CNY 15 per 1M tokens for Seedance 1.0 Pro). ERNIE Bot has been free for individual users since March 2025, replacing “Ernie Bot subscription”; ERNIE 5 and ERNIE X1.1 also show their Baidu Qianfan API prices (CNY 6/24 and CNY 1/4 per 1M input/output tokens)
• Adobe Podcast — Free plan, with premium features in the Adobe Firefly Pro plan at $19.99 a month (was “Adobe Creative Cloud subscription”)

1,176 models · 329 vendors · 38 fields corrected on 20 models

Corrected

• Amazon — Alexa+ is free with Prime or $19.99 a month, with a free usage-limited chat tier at Alexa.com and in the Alexa app. Nova Micro, Lite and Pro now show their Amazon Bedrock prices ($0.035/$0.14, $0.06/$0.24 and $0.80/$3.20 per 1M input/output tokens); the blank Nova 2 Lite ($0.30/$2.50), Nova 2 Omni preview and Nova 2 Sonic prices are filled in; Titan Embeddings shows its per-token prices. Amazon Q Developer is Free or Pro at $19 per user a month, and CodeWhisperer is now part of it. Titan Image Generator is retired: its last version reached end of life on Bedrock on June 30, 2026. These rows had said “Alexa/AWS services; consumer pricing varies”
• AlexaTM — AlexaTM is an internal Alexa research model. AlexaTM 20B weights are free for non-commercial use under the Alexa Teacher Model License through SageMaker JumpStart, so its Open Source, Deployment and Free Tier fields now match
• Apple — Apple Intelligence is free on compatible devices, with daily limits on features that run on Apple’s servers. The on-device Apple Foundation Model is free for developers to use, and Private Cloud Compute is free for small developers. MLX is open source (MIT); DiffuCoder and OpenELM weights are for research only; Ferret’s weight deltas are CC BY-NC. These rows had said “Included in Apple ecosystem; developer/research models vary”

1,176 models · 329 vendors · 51 fields corrected on 28 models

Corrected

• IBM and AI21 — IBM Granite 3.1, 3.2, 3.3 and 4.0 Tiny, Granite Code and Bamba-9B v2 showed “IBM watsonx pricing varies”; they are free open weights under Apache 2.0. Jamba is Apache 2.0 and Jamba 1.5 is under the Jamba Open Model License
• Cohere — Command A, Command R, Command R+, Command R7B, Aya 23 and Aya Expanse showed “Cohere API pricing”. Their weights are for non-commercial use only (CC BY-NC 4.0); Command R7B ($0.0375/$0.15 per 1M tokens) and Aya Expanse ($0.50/$1.50) now show API prices, and Command R and R+ their legacy prices for existing customers
• NVIDIA and Zhipu — Nemotron Ultra, Super and Nano, Nemotron-4 and Cosmos 1.0 are under the NVIDIA Open Model License; Instant-NGP is research-only source code. GLM-4-0414 and GLM-4.5 are MIT; CogVideoX-2B is Apache 2.0 and CogVideoX-5B the CogVideoX License; ChatGLM-6B has its own license; CogVLM and CogVLM2 allow free commercial use after registration

1,176 models · 329 vendors · 48 fields corrected on 24 models

Corrected

• Stability AI licenses — 13 Stability AI rows said only “Licensing/API varies”. Stable Diffusion 3 Medium, Stable Audio Open, Stable LM 2, SDXL Turbo and SV3D are Core Models under the Stability AI Community License (free, including commercial use for individuals and companies under $1M in annual revenue); Stable Cascade and Stable Zero123 are non-commercial only; SDXL is CreativeML Open RAIL++-M; StableCode and Japanese StableLM base are Apache 2.0; StableLM Alpha base models are CC BY-SA 4.0 and tuned models CC BY-NC-SA 4.0
• Stable Audio and Clipdrop plans — Stable Audio now shows its plans (Free with 50 credits a month for non-commercial use; Solo $12, Session $30, Producer $90 and Studio $199 a month) and Clipdrop its plans (Free; Pro $15 a month; usage-based API)
• Hugging Face models — 11 Hugging Face model rows showed Hugging Face’s platform plans instead of the model’s terms. They now show their licenses: Apache 2.0 for Parler-TTS, DistilBERT, Idefics2, Idefics3, SmolLM, SmolLM2 and SmolVLM; MIT for Zephyr 7B; BigScience BLOOM RAIL 1.0 for BLOOM and BLOOMZ

1,176 models · 329 vendors · 71 fields corrected on 37 models

Corrected

• Open-weight Qwen models — 29 open-weight Qwen-family rows had vague pricing (mostly just “Alibaba Cloud Model Studio”). They now show that the weights are free and under which license: Apache 2.0 for Qwen3, QwQ-32B, Qwen3-Omni, Qwen3-Next, Qwen3.5, QwenLong-L1, ParScale, PolyLM and most Qwen2/2.5 sizes; the Qwen license for Qwen2.5-72B; the Tongyi Qianwen license for Qwen 1, Qwen1.5, Qwen-VL, Qwen-Audio and Qwen2-72B; the SeaLLM license for Babel-83B; MIT for GTE-Large; CC BY-NC 4.0 (non-commercial) for ModelScope T2V
• Current Qwen and Wan API prices — Qwen3.6 Plus, Qwen3.6-35B-A3B, Qwen3.7-Max, Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen3.8-Max and Wan 3.0 now show Alibaba Cloud Model Studio’s current international list prices. Qwen3.8-Max and Qwen3.8-Flash-Next weights are now public (Qwen3.8-Max License; Qwen Community License 1.0)
• Description — Qwen2.5-Max was described as an open-source family in multiple sizes; it is a single large mixture-of-experts model offered through Alibaba Cloud’s API and Qwen Chat

1,176 models · 329 vendors · 58 fields corrected on 30 models

Corrected

• xAI API retirements — Grok 3, Grok 4, Grok 4 Fast, Grok 4.1 Fast and Grok Code Fast 1 are marked retired: xAI removed them from its API on May 15, 2026 and now routes their requests to Grok 4.3 (Grok Code Fast 1 to grok-build-0.1). Several of these rows showed the X Premium+ plan price
• Grok prices and open weights — Grok 4.3 and Grok 4.20 now show their API prices from xAI’s pricing page ($1.25/M input, $2.50/M output; Grok 4.20 previously carried an unsourced estimate). Grok-1 (Apache 2.0) and Grok 2 (xAI Community License) are marked as free open weights; Grok 2 showed X Premium+ at $16/mo and no free tier
• Mistral API retirements — Magistral Medium, Devstral Medium, Mistral Medium 3, Mistral Saba, Mistral Small 3 and 3.1, Mistral Large 2, Pixtral Large, Pixtral 12B and Devstral 2 are marked retired from Mistral’s API with their retirement dates; several showed the Le Chat Pro price. Open-weight versions note their licenses (Mistral Large 2 and Pixtral Large are research-only under the Mistral Research License)
• Mistral prices — Mistral Large 3 ($0.50/$1.50 per 1M tokens), Mistral Embed ($0.10), Ministral 3 ($0.10–$0.20) and Voxtral (Voxtral Small $0.004/min; Voxtral Mini retired) now show current API prices from Mistral’s model pages
• Descriptions — Magistral Medium and Mistral Medium 3 carried a biomedical-model template that did not describe them; Aurora (xAI) described Microsoft’s Aurora weather model. All three now describe the right model from official announcements

1,176 models · 329 vendors · 67 fields corrected on 41 models

Corrected

• ChatGPT prices — ChatGPT, ChatGPT Canvas and ChatGPT Images 2.0 showed ChatGPT Pro at $200/mo; Pro now starts at $100/mo (Free, Go $8 and Plus $20 unchanged). GPT-Live now shows how it is included in ChatGPT Voice
• Models no longer in ChatGPT — GPT-5, GPT-5.2, GPT-5.4, GPT-5.5 and GPT-4.1 were listed with ChatGPT plan prices and GPT-4o mini as the free-plan default, but ChatGPT now offers GPT-6 and GPT-5.6 models. They now show their API prices from OpenAI’s model pages, with shutdown dates where OpenAI has set one
• Retired and discontinued — GPT-4 Vision preview, DALL·E 2, DALL·E 3, o1-preview, o1-mini and GPT-4.5 Preview are marked retired with their API shutdown dates; Sora and Sora 2 are marked discontinued (app closed April 26, 2026; API shut down Sept 24, 2026); GPT-4b micro is a research model that OpenAI says is not broadly available, so its free tier changed to N/A
• API shutdown dates — 22 deprecated OpenAI models now show when their API access ends, from Oct 23, 2026 (GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, o1, o3-mini, o4-mini, GPT-4.1 nano, GPT Image 1) to Apr 1, 2027 (GPT-5.1, GPT-5.3-Codex, GPT-5.4 nano)

1,176 models · 329 vendors · 66 fields corrected on 34 models

Corrected

• Retired Claude models no longer shown as for sale — 18 dossiers for Claude models that Anthropic has retired (Claude 1, Claude Instant, Claude 2 and 2.1, the Claude 3, 3.5 and 3.7 models, Claude Sonnet 4, Opus 4 and Opus 4.1) still listed the current Claude app plan prices. They now give the date each model was retired from the Claude API, from Anthropic’s model deprecations page
• Claude API prices — Claude Opus 4.5 ($5/M input tokens), Sonnet 4.5 (deprecated, retiring Nov 30, 2026), Sonnet 4.6 and Haiku 4.5 now show their API prices instead of the app plan list
• Claude in Chrome and Claude Code — Both now say they come with Claude Pro ($20/mo, or $17/mo billed annually) or Max (from $100/mo), not the free plan; Claude in Chrome’s free tier changed from yes to no
• DeepSeek — Ten DeepSeek dossiers shared an old “$0.27/M input” API price that DeepSeek no longer charges. Nine open-weights models (DeepSeek-V3.1, the R1 distills, VL, VL2, V2, Coder, Coder-V2-Lite and LLM) now show their Hugging Face license, and DeepSeek Chat shows the free app plus the current DeepSeek-V4.1-Flash API price

1,176 models · 329 vendors · 70 fields corrected on 39 models

Corrected

• Products wrongly priced as a Google AI Pro subscription — Eight dossiers said “Google AI Pro (Gemini): $19.99/mo” for products that are not sold that way. They now show each product’s real cost from official pages: Google Translate (free; Cloud Translation API $20 per million characters), Photomath (Plus $9.99/mo), Socratic (free; its site now redirects to Google Lens), Deep Dream Generator (an independent site, Basic $9/mo), AI Overviews (free in Search), AlphaFold 3 (free for non-commercial use), Gemini for Workspace (Business Starter $7 per user/mo) and Google Assistant (free; paid plans from $4.99/mo)
• Retired Google models no longer shown as for sale — PaLM 2, LaMDA, Bard, Gemini 1.0, 1.0 Ultra, 1.5, 1.5 Pro, 1.5 Flash, 2.0, 2.0 Flash exp, 2.0 Pro, the 2.5 Flash and 2.5 Pro previews, Robotics-ER 1.5 and Robotics-ER 1.6 now say when Google retired them, citing Google Cloud’s model lifecycle page and the Gemini API deprecations page
• Current Google prices — Gemini 2.5 Pro ($1.25/M input tokens), 2.5 Flash and 3.1 Flash-Lite now show Gemini API prices; the Gemini app, Canvas and Flow list the free plan plus Google AI Plus ($4.99/mo), Pro ($19.99/mo) and Ultra (from $99.99/mo); Deep Think moves to Google AI Ultra; Flow, Veo 2 and Veo 3 drop the old $249.99/mo Ultra price (Veo 3 was retired June 30, 2026); Imagen 3, Imagen 4, Lyria 2 and Lyria 3 show Google Cloud API prices, since ImageFX and MusicFX now redirect to Google Flow and Flow Music
• Flow free tier — Google’s subscriptions page includes Flow with limited access in the free plan, so Flow’s free tier changed from no to yes (limited)

1,176 models · 329 vendors · 26 fields corrected on 25 models

Corrected

• Prices filled for 25 more models — 25 dossiers that showed no cost now list prices or license terms from official sources: API rates for Grok Imagine Video 1.5 ($0.08/sec at 480p), MAI-Thinking-1 ($2/M input tokens), MAI-Cyber-1-Flash, MAI-Image-2.5, Gemini 3.5 Flash-Lite ($0.30/M input tokens), Lyria 3.5 ($0.08 per song), GLM-5.2 and Kimi K2.7 Code; license terms for open-weights models (Holo 3.1, Inkling, Ling-3.0-Flash, Jais 30B, MiniMax H3, Ideogram 4.0, LTX-Video) and open-source projects (Aider, Continue, DiffusionBee, ULMFiT, Transformer-XL, Gaussian Splatting, Koala); and plans for v0 (Plus $30 per user/mo), NovelAI (from $10/mo) and LM Studio (free; Bionic+ $20/mo)

1,176 models · 329 vendors · 13 fields corrected on 13 models

Corrected

• Prices filled for 13 more models — 13 dossiers that showed no cost now list the plans and prices on each vendor’s own page: Unitree G1 (from US$13,500), Runway Gen-4.5 (Standard $15/mo), Luma Ray3 (Plus $30/mo), Krea (Basic $9/mo), Julius (Plus $20/mo), D-ID (Lite $4.70/mo billed annually), Draw Things (Draw Things+ $8.99/mo), n8n (Cloud Starter €20/mo), Cleo (Plus $5.99/mo), Mistral OCR 3 ($2 per 1,000 pages), Base 1 (Base44 plans), Langflow (free, open source) and Nomic Embed (free open weights)

1,176 models · 329 vendors · 26 fields corrected on 25 models

Corrected

• Prices filled for 25 products — 25 dossiers that showed no cost now list the plans and prices on each vendor’s own pricing page, among them HeyGen (Creator $29/mo), Warp (Build $20/mo), Kagi (from $5/mo), Figma (Professional seat $16/mo), Devin (Pro $20/mo), Wolfram Alpha (Pro $9.99/mo), Brave Leo (Premium $14.99/mo), Khanmigo ($4/mo, free for teachers) and Opera Aria (free)
• Devin free tier — Devin’s pricing page now lists a Free $0 plan with a light quota, so its free tier changed from no to yes (limited)

1,176 models · 329 vendors · 40 fields corrected on 40 models

Corrected

• Cost filled for open-weights models — 36 dossiers that showed no cost now say Free (open weights), with the license named on the model’s own Hugging Face or GitHub page: among them Gemma 4 12B, Mellum 2, Allegro and Mochi 1 (Apache 2.0), LongCat-Flash-Lite-Sparse and Step-Video-T2V (MIT), Laguna S and XS 2.1 and Nemotron 3 Ultra (OpenMDW-1.1), LFM2.5 2.6B (LFM Open License v1.0), MARS5 and YOLOv10 (AGPL-3.0), and Zeroscope v2 (CC BY-NC 4.0, non-commercial)
• Research models marked as not sold — AlexNet, the feedforward neural language model, RWTH Aachen’s LSTM language model and the RNN language model now read N/A (research), citing their papers

1,176 models · 329 vendors · 13 fields corrected on 11 models

Corrected

• Free tier filled — 9 of the 15 models that showed no free-tier answer now have one, each taken from an official page: Claude Opus 5.5 (no — Opus is not on Claude’s Free plan), GPT-5.5 Pro, GPT-5 mini, GPT-5 nano and GPT-5.4 nano (no — OpenAI lists its free API tier as not supported for them), Midjourney V1 Video and V8.1 (no — every Midjourney plan is a paid subscription), and FLUX.1 Kontext and FLUX.2 (yes — their [dev] weights are free to download under non-commercial licenses). MiniMax-M3 changed from unknown to yes for its free open weights
• Pricing and deployment fixes — GPT-5.4 nano now shows its full API price ($0.20/M input, $1.25/M output, $0.02/M cached input), and GPT-J-6B is labeled as a local open-weights model instead of Cloud / API / Mobile

1,176 models · 329 vendors · 64 fields corrected on 64 models

Corrected

• X handles filled — 64 more dossiers now list an X account, each taken from the company’s or project’s own website, help center, page metadata or official repository, with the exact spelling checked: among them Canva (@canva), Notion (@NotionHQ), Duolingo (@duolingo), Salesforce (@salesforce), Snapchat (@Snapchat), LinkedIn (@LinkedIn), Leonardo.Ai (@LeonardoAi_), Lovable (@Lovable), ComfyUI (@ComfyUI), Open WebUI (@OpenWebUI), Vicuna (@lmsysorg) and the StarCoder models (@BigCodeProject). The two Google Brain ViT models now point to @GoogleDeepMind, the lab Google Brain became part of in 2023

1,176 models · 329 vendors · 106 fields corrected on 106 models

Corrected

• X handles filled — 106 dossiers that showed no X account now list one, each taken from the X link on the company’s own website with the exact spelling checked: among them Figma (@figma), Framer (@framer), DeepL (@DeepLcom), Descript (@descript), HeyGen (@HeyGen), Groq (@GroqLLC), Harvey (@harvey), Slack (@slackhq), v0 (@v0), Zoom (@Zoom), Cognition’s Devin (@cognition), Elai (@elai__io) and D-ID (@D_ID_). Sites that link no account, more than one candidate, or only a renamed parent company were left blank

1,176 models · 329 vendors · 38 fields corrected on 17 models

Corrected

• More research models no longer listed as cloud products — 13 models that were only published as research — Genie 2, NaturalSpeech 3, RT-2, CM3Leon, Voicebox, VALL-E, VALL-E X, PaLM-E, Make-A-Video, the original DALL·E, AlphaGo, AlphaZero and the Large Concept Model — claimed a cloud deployment and a free tier. They now read Research, with no free tier; Voicebox’s and Make-A-Video’s own pages say they were withheld from public release
• Open weights labeled correctly — Kosmos-2 (MIT) and Pixtral 12B (Apache 2.0) now show open weights. Pixtral Large (Mistral Research License) and Command A (CC-BY-NC 4.0) publish weights for non-commercial use only and are marked Partial, like Command R
• Microsoft papers as primary sources — NaturalSpeech 3, VALL-E, VALL-E X and Kosmos-2 linked to Microsoft Research’s home page; they now link their papers on arXiv

1,176 models · 329 vendors · 16 fields corrected on 8 models

Corrected

• Open weights released after launch — 8 models whose weights reached Hugging Face after their dossiers were written still said closed, pending or promised. MiniMax-M3, DeepSeek-V4-Flash-Vision-Exp, GLM-5.3, Qwen3.8-Max (published as Qwen3.8-2.4T-A95B), Qwen3.8-Flash-Next, Ling-3.0-Flash and Grok 2 now show open weights with local deployment alongside their APIs; Mistral Large 2 (2407), whose weights carry the research-only Mistral Research License, is marked Partial

1,176 models · 329 vendors · 40 fields corrected on 27 models

Corrected

• Research models no longer listed as cloud products — 12 models that were only ever published as research — Chinchilla, Gopher, Gato, MuZero, the 2017 Transformer (and its duplicate entry), Seq2Seq LSTM, Minerva, GLaM, FLAN 137B, CoAtNet and AlexNet — were listed as Cloud only (proprietary). They now read Research, and the seven that claimed a free tier no longer do
• Google papers as primary sources — 21 Google dossiers linked their primary source to the generic Gemini API models page; they now link the original paper on arXiv, among them PaLM, PaLM 2, LaMDA, T5, BERT, XLNet, Inception v3, Xception, EfficientNetV2, ViT-22B and word2vec

1,176 models · 329 vendors · 30 fields corrected on 15 models

Corrected

• Open weights mislabeled as proprietary — 15 models whose weights are public on Hugging Face were listed as proprietary and cloud-only. Grok-1, GLM-4-32B-0414, K2-Think, s1-32B, Sky-T1-32B-Preview, UI-TARS-72B, Comma v0.1-2T, YuLan-Mini, ParScale, Babel-83B, EVA, wav2vec 2.0 and Fugaku-LLM now show open weights and local deployment (plus hosted access where a provider serves them). Apple DiffuCoder and Meta BLT, whose licenses allow only non-commercial research, are marked Partial with research-license local deployment, as Chameleon already was

1,176 models · 329 vendors · 33 fields corrected on 33 models

Corrected

• Academic releases marked — 33 university research releases that had a blank CEO field now read N/A (academic research), matching how the database already treats Stanford Alpaca, ColBERTv2 and GloVe. Each is backed by the paper or project page naming the releasing lab: AlexNet (University of Toronto), VGG16 and VGG19 (Oxford), YOLOv3 (University of Washington), YOLOv10 (Tsinghua), Vicuna (LMSYS), Koala, Starling and Sky-T1 (UC Berkeley), ControlNet (Stanford), LLaVA (UW–Madison), Apertus (EPFL, ETH Zurich and CSCS), Meditron (EPFL), OpenFold (Columbia), RFdiffusion (UW Institute for Protein Design) and others

1,176 models · 329 vendors · 9 fields corrected on 7 models

Corrected

• Missing CEOs filled — 5 more company dossiers now name a CEO: WellSaid (Benjamin Dorr, CEO since May 2025), Topaz Labs (Eric Yang, who stays on after Adobe’s agreement to acquire the company), Loudly (Rory Kenny), Julius (Rahul Sonwalkar) and Kapwing (Julia Enthoven)
• More key staffers named — Topaz Photo AI listed a generic team; it now names CEO Eric Yang. Loudly, LALAL.AI and Cleanvoice had no key staffers; they now list co-founder and CEO Rory Kenny, co-founder Nik Pogorsky and founder Adrian Spataru

1,176 models · 329 vendors · 9 fields corrected on 9 models

Corrected

• Missing CEOs filled — 9 company dossiers that had no CEO now name one: Genmo (Paras Jain), Mem (Kevin Moody), Pictory (Vikram Chalana), Beatoven.ai (Mansoor Rahimat Khan), Fish Audio (Rissa Cao), Fractal Analytics (Group Chief Executive Srikanth Velamakanni), Beautiful.ai (Jason Lapp), Perceptron (Armen Aghajanyan) and Pleias (Anastasia Stasenko)

1,176 models · 329 vendors · 8 fields corrected on 8 models

Corrected

• More generic “team” entries replaced — 8 more dossiers now name people instead of a generic team: ByteDance’s Seedance (Seed head Wu Yonghui), Shanghai AI Lab’s InternLM 2.5 and Agents-A1 (Director and Chief Scientist Bowen Zhou), Vicuna and Vicuna-13B (corresponding authors Lianmin Zheng and Hao Zhang, with advisors Joseph Gonzalez, Ion Stoica and Eric Xing), RFdiffusion (its equal-contribution authors and David Baker), SambaNova’s EvaByte and Ultralytics’ YOLOv5

1,176 models · 329 vendors · 55 fields corrected on 54 models

Corrected

• Generic “team” entries replaced with named people — 54 dossiers whose Key Staffers relied on a generic placeholder such as “MiniMax team”, “LG AI Research team”, “Stanford University AI research” or “Non-profit, French government-backed” now name real people. MiniMax’s seven models list its LLM and visual-model research leaders and COO from its investor-relations page; Kling’s five list division head Gai Kun; LG’s EXAONE models list co-presidents Honglak Lee and Woohyung Lim; Kyutai, Udio, Sesame, Lensa, WellSaid Labs and TII’s Falcon list their leaders or founders
• Research models credit their authors — Paper authors now appear for s1, Alpaca, ColBERTv2, GloVe, YuLan-Mini, mmBERT, Apertus, VGG16 and VGG19, YOLOv3, YOLOv8, Spectra 1.1, OpenAssistant, ModernBERT, Nucleotide Transformer, OLMo, OLMo 2, Tülu 3, WildGuard, FlexOlmo, SciBERT, ELMo, jina-embeddings-v3, KAT-V1, Qwerky-72B, Fugaku-LLM and the RNN and LSTM language models. TII’s Falcon dossier also gained its CEO, Najwa Aaraj

1,176 models · 329 vendors · 10 fields corrected on 10 models

Corrected

• More key staffers named — 10 more dossiers now list key people beyond the CEO: the founders of Reve (2.0 and 2.1), Glif, Colossyan and Cleo, Quizlet founder Andrew Sutherland, Tilde co-founder Andrejs Vasiļjevs, Ant Group CTO He Zhengyu for InclusionAI’s Ring-1T-preview, Kunlun Tech chairman and CEO Fang Han for SkyReels, and Poe product lead Gareth Jones

1,176 models · 329 vendors · 40 fields corrected on 35 models

Corrected

• Missing CEOs filled and two placeholders replaced — Vidu’s maker Shengshu has been led by CEO Luo Yihang since March 2025, with co-founder Tang Jiayu now president; Camb.AI (Avneesh Prakash), Invoke (Kent Keirsey) and Boomy (Alex Mitchell) also gained CEOs. Moondream and Fuyu-8B listed a generic team instead of a chief executive; they now name Jay Allen of M87 Labs and Zach Brock, who took over Adept in June 2024
• More key staffers named — 34 more dossiers now list key people beyond the CEO, from SEC filings, official leadership and newsroom pages, model cards and tier-1 press: Figma, Wix, monday.com, Stripe, Slack, Mailchimp, Vercel v0, Jasper, Leonardo, Descript, Replika, Tabnine, Tome, Sudowrite, Make, Moises, Base44, remove.bg, iFlytek SparkDesk and Xiaomi MiMo among them, plus the authors of Fuyu-8B and Jais 30B

1,176 models · 329 vendors · 26 fields corrected on 24 models

Corrected

• Groq and Mailchimp leaders corrected — Groq’s June 22, 2026 funding release says it is now led by CEO Adam Winter (we listed Simon Edwards, named CEO when Nvidia licensed Groq’s technology in December 2025), and Intuit named Matt Idema Senior Vice President of Mailchimp in February 2025 (we listed Rania Succar). Groq’s dossier also lists its new CTO, CPO and COO
• More key staffers named — 21 more dossiers now list key people beyond the CEO, including HeyGen, Granola, Elicit, Otter.ai, Framer, Superhuman, Sourcegraph Cody, tl;dv, Murf, Squarespace, Loom, Copy.ai, Tesla Optimus, Magnific and Pictory, plus the authors of Playground v2.5, MOSS-VL-Realtime and ELYZA’s Japanese Llama 2. Tilde, Typecast and Magnific also gained CEOs

1,176 models · 329 vendors · 48 fields corrected on 41 models

Corrected

• More key staffers named — 35 more dossiers now list key people beyond the CEO, from official leadership, investor-relations and team pages and tier-1 press: DeepL, Duolingo Max, Zoom AI Companion, Snapchat My AI, Intercom Fin, Gamma, Grammarly, Harvey, Speechify, Fireflies, You.com, Zendesk AI, CoCounsel, Krisp, Lovable, Zapier and Arc Search among them. Duolingo’s investor page shows co-founder Severin Hacker is no longer CTO, so Duolingo Max lists its current engineering and product chiefs
• Missing CEOs filled — 13 dossiers that had no CEO now name one: Hex (Barry McCardel), Bolt.new (Eric Simons), Krea (Victor Perez), Kaiber (Victor Wang), Viggle (Hang Chu), Recall.ai (David Gu), Resemble AI (Zohaib Ahmed), Dify (Luyu Zhang), Khanmigo (Sal Khan), Meshy (Ethan Yuanming Hu), Moises (Geraldo Ramos), Podcastle (Arto Yeritsyan) and Elai (Vitalii Romanchenko)

1,176 models · 329 vendors · 59 fields corrected on 54 models

Corrected

• Nomic AI and QuillBot CEOs corrected — Nomic’s own About page says founder Andriy Mulyar has been CEO since 2024 (we listed Brandon Duderstadt), and QuillBot’s lists Andrew Grauer as CEO (we listed co-founder Rohan Gupta). Political notes written about the former CEOs were removed
• More key staffers named — 54 more dossiers now list key people beyond the CEO, sourced from official leadership pages, an SEC prospectus and tier-1 press: Canva, Notion, Asana, Salesforce, CrowdStrike, Cognition’s Devin and Windsurf, ClickUp, Remini (Bending Spoons), Opera Aria, Discord, DataRobot, Riffusion (now Google Flow Music) and Sber’s GigaChat among them. Research models including TripoSR, Allegro, SaulLM-7B, MAP-Neo, mGPT, Kandinsky 3.0, SOLAR 10.7B, Fish-Speech and TNG’s Chimera models now list their lead authors

1,176 models · 329 vendors · 49 fields corrected on 49 models

Corrected

• Zhipu AI leadership corrected — eleven Zhipu dossiers called Tang Jie the company’s CEO. Per Caixin Global’s July 2026 profile, Zhang Peng is CEO, Tang Jie is Chief Scientist and Liu Debing is Chairman; all fourteen Zhipu models, from ChatGLM-6B to GLM-5.3, now list the three
• More key staffers named — EleutherAI’s Pythia, GPT-NeoX, GPT-Neo and Comma now list their lead authors; research models including BLIP-2, WizardCoder, YOLOv10, Grounding DINO, DenseNet, Meditron, ColPali, F5-TTS, PhotoMaker, Koala, Starling, Sailor2 and RWKV-7 list lead and senior authors from their papers; Open WebUI, Fooocus, Tortoise TTS, Kokoro, Saiga and Continue list their creators; Brave and HubSpot list their technical leaders

1,176 models · 329 vendors · 73 fields corrected on 71 models

Corrected

• Key staffers named — 58 dossiers that showed no key staffers or only a generic “research/maintainer team” now name the people behind the model, from the model’s paper, the maker’s own pages or tier-1 press. Shopify Sidekick now lists CTO Mikhail Parakhin, VP of Product Vanessa Lee and Andrew McNamara, the Director of Applied ML who leads Sidekick; research models such as StarCoder, BGE, LLaVA, ControlNet, Depth Anything and Gaussian Splatting list their lead and senior authors
• MiniMax X handle — MiniMax’s official X account is @MiniMax_AI; the old @MiniMax__AI handle does not exist. Hailuo rows now point to MiniMax’s video account @Hailuo_AI and Talkie to @talkie_ai
• CEOs — Udio now lists co-founder and CEO Andrew Sanchez, Civitai Justin Maier, ComfyUI Yoland Yan, and SEA-LION AI Singapore’s Executive Chairman Christian Wolfrum

1,176 models · 329 vendors · 55 fields corrected on 48 models

Corrected

• Real prices replace placeholders — 47 dossiers that said “See official pricing page” or “verify current regional pricing” now list actual plans and prices from each vendor’s own pricing page or help center, including Canva, Grammarly, Slack, Zoom, Shopify, HubSpot, Mailchimp, monday.com, Descript, Lovable, Replit, Make and Consensus
• Shutdowns and rebrands — remove.bg’s standalone site closes December 1, 2026 as it moves into Canva; Obviously AI is now Zams and obviously.ai is an archive; tabnine.com now redirects to Tricentis; Podcastle now runs as Async; and Sourcegraph ended Cody Free and Pro on July 23, 2025, so Cody’s free tier now reads No
• Still placeholders — five rows keep the pointer to the official page because their prices could not be verified: Beatoven.ai (pricing page returns 404), Fliki (prices load only in the app), Loudly (prices did not render), Moises (prices behind a login) and Stripe Copilot (no Stripe product by that name could be confirmed)

1,176 models · 329 vendors · 40 fields corrected on 40 models

Corrected

• Blank jurisdictions filled — 40 more dossiers showed [UNKNOWN] for jurisdiction; each now uses the country named in the company’s own privacy policy, terms or imprint, its SEC filer record, the model’s paper, or press naming its base. For example Descript, Gamma, Notion, Ollama and Warp are in the USA, Framer in the Netherlands, TextCortex in Germany, Soundraw and ELYZA in Japan, Haiper in the UK, Make in the Czech Republic and Vidu in China; PixVerse lists Singapore and China, Riverside the USA and Israel, and Invideo’s terms name entities in Singapore, the USA and India
• Research models — ULMFiT now shows the USA and Ireland (fast.ai and NUI Galway), LLM360’s Amber the UAE and USA (MBZUAI, Petuum and Cerebras), TripoSR the UK and China (Stability AI with Tripo), and both Depth Anything models China (HKU and TikTok)
• Still unknown — ten jurisdictions stay blank because no reliable source names them: Craiyon, SaulLM-7B, Jan, Kaiber, Mem, both Dia models, NightCafe, Play.ht and Sudowrite

1,176 models · 329 vendors · 47 fields corrected on 35 models

Corrected

• Primary sources — 16 dossiers had no Primary Source link or only a generic one; they now link to the official announcement or model page, including Qwen3.7-Max, Muse Spark, Gemma 4, GPT-5.3-Codex, Gemini 3 Flash, Nano Banana and Nano Banana Pro, Veo 3.1, Kimi K2.5 and K2 Thinking, and MiniMax M2, M2.5 and M2.7
• Descriptions — 11 dossiers showed [UNKNOWN] under What Makes It Special; they now describe the model from its model card, paper or launch coverage, for example Tencent’s Hunyuan T1 and Turbo S, Perplexity’s R1 1776, LinkedIn’s EON-8B, TeleAI’s T1 and the early neural language models from the Université de Montréal, Brno University of Technology with Johns Hopkins, and RWTH Aachen
• Tome and Play.ht — both products are gone: Tome’s founders shut the app down to build Lightfield, and Play.ht shut down after Meta acquired PlayAI in July 2025; their dossiers now say so
• Open weights — TeleAI’s T1, Featherless AI’s Qwerky-72B (now QRWKV-72B) and Perplexity’s R1 1776 all published their weights but were marked proprietary; they now show open weights
• Cerebras 1T — the trillion-parameter model Cerebras trained with Sandia National Laboratories was a NeurIPS 2024 training demonstration, not an API product; its pricing and deployment now say so
• Statuses and deployment — blank stock statuses filled for SkyReels V1 (Kunlun Tech), LTX-Video, Pyramid Flow, GPT-J-6B, Whisper.cpp and GPT4All, and blank deployment types for PolyLM and Loudly AI

1,176 models · 329 vendors · 52 fields corrected on 45 models

Corrected

• CEO filled in — 27 more dossiers showed [UNKNOWN] for CEO. Most now come from the company’s own pages or press releases (for example Zeb Evans at ClickUp, Joshua Xu at HeyGen, Chris Pedregal at Granola, Vladimir Prelovac at Kagi and Davit Baghdasaryan at Krisp) or from major press (for example Wade Foster at Zapier, Winston Weinberg at Harvey, Jeff Morgan at Ollama, and Daniel Vassilev and Jacky Koh, Relevance AI’s co-CEOs); Laura Burkhauser has led Descript since founder Andrew Mason stepped aside in September 2025; the rest use the CEO named on Wikipedia (Anthony Casalena at Squarespace, Kurt Beidler at Quizlet, Elon Musk at Tesla and Wang Xingxing at Unitree)
• Acquired and parent companies — 10 dossiers for products whose maker was acquired or belongs to a larger group now show the parent’s CEO: Mike Cannon-Brookes (Atlassian) for Loom AI and Arc Search, Scott Wu (Cognition) for Codeium, Ryan Westwood (Fullcast) for Copy.ai, Melanie Perkins (Canva) for Remove.bg, Herman Gref (Sber) for mGPT, Cyril Han (Ant Group) for InclusionAI Ring-1T-preview, Mark Zuckerberg (Meta) for Play.ht, Sundar Pichai (Google) for Riffusion and Arkady Volozh (Nebius) for Tavily
• Copy.ai and Remove.bg — Fullcast acquired Copy.ai in October 2025, and Canva acquired Remove.bg maker Kaleido in February 2021; both statuses now say so
• Shutterstock — Paul Hennessy stepped down as CEO on July 13, 2026, and CFO Rik Powell became interim CEO; Amper Music’s dossier now shows him
• CEO politics — 12 dossiers now carry the same politics note as their CEO’s other dossiers: Tesla Optimus Gen 2, Play.ht, Riffusion, mGPT and Ring-1T-preview, which had a placeholder; Windsurf SWE-1, which was blank; and four Meta AI and two Google Brain models that had older or different wording

1,176 models · 329 vendors · 5 fields corrected on 2 models

Corrected

• Riffusion — the Riffusion team rebranded as ProducerAI, which Google acquired in February 2026, and riffusion.com now serves Google Flow Music; its status, description and pricing now say so
• Discord Clyde — Discord shut its experimental Clyde chatbot down on December 1, 2023; its dossier now says so

1,176 models · 329 vendors · 12 fields corrected on 12 models

Corrected

• CEO filled in — 12 more dossiers showed [UNKNOWN] for CEO. Discord’s own press release names Humam Sakhnini (CEO since April 2025); company pages name Krish Ramineni (Fireflies), Nadav Keyson (Riverside), Samanyou Garg (Writesonic), Jennifer Smith (Scribe) and Andreas Stuhlmüller (Elicit); a Stripe newsroom release quotes Patrick Collison as cofounder and CEO; London Tech Week’s 2026 speaker profile lists Ivan Zhao as Notion’s co-founder and CEO; and Wikipedia infoboxes give Roy Mann and Eran Zinman (Monday.com co-CEOs), Choi Soo-yeon (Naver) and Rania Succar (Mailchimp)

1,176 models · 329 vendors · 44 fields corrected on 43 models

Corrected

• CEO filled in — 43 dossiers showed [UNKNOWN] for CEO. Most now come from the company’s own team, about or press page (for example Eoghan McCabe at Intercom, Yamini Rangan at HubSpot, Timothy Young at Jasper, Will Bryk at Exa, Richard Socher at You.com and Eric Yuan at Zoom); the rest use the CEO named in the company’s Wikipedia infobox (for example Dan Rogers at Asana, Luis von Ahn at Duolingo, Tobias Lütke at Shopify and Evan Spiegel at Snap)
• Tabnine — Tricentis acquired Tabnine in July 2026; its status now says so, and its CEO is founder Dror Weiss

1,176 models · 329 vendors · 30 fields corrected on 29 models

Corrected

• Pricing placeholders replaced — 25 dossiers showed “See official pricing page” instead of a price; they now list the plans and prices from each company’s own pricing page, with the billing basis where the page states it (for example Windsurf, DeepL, Fireflies, Notion, Loom, Zapier, Synthesia, Squarespace, Asana and Intercom’s Fin)
• Open-weight models — BioMistral and Replit Code v1.5 were also on the placeholder; as open-weight releases they now read “Free (open weights)”
• Synthesia — its main dossier said “No free tier”, but Synthesia’s pricing page lists a free Basic plan
• Grammarly and Superhuman — both dossiers had no CEO; Shishir Mehrotra has led the company, which took the Superhuman name in 2025, since Grammarly acquired Coda

1,176 models · 329 vendors · 20 fields corrected on 20 models

Corrected

• Blank jurisdictions filled — 19 more dossiers showed [UNKNOWN] for jurisdiction; each now uses the country named on the company’s own privacy policy, terms or imprint (for example Grammarly, Copy.ai, Fireflies, Krea and Murf in the USA, Relevance AI and Splash in Australia, Pixlr in Singapore, Typecast maker Neosapience in South Korea, and Granola, which lists both a US and a UK company)
• Pixtral Large — Mistral now marks Pixtral Large as deprecated and no longer maintained; its pricing notes this

1,176 models · 329 vendors · 30 fields corrected on 30 models

Corrected

• Blank jurisdictions filled — 30 more dossiers showed [UNKNOWN] for jurisdiction. Most now come from the company’s own legal pages, which name the operating entity and its address (for example Jasper and Taskade in the USA, Photoroom in France, Opera in Norway, Canva Austria, which runs remove.bg, in Austria, Lalal.ai in Switzerland, Loudly and tl;dv in Germany, and Viggle in Canada); Snap, Stripe, Squarespace, Otter.ai, QuillBot, Harvey and Loom use the headquarters in their Wikipedia infobox

1,176 models · 329 vendors · 109 fields corrected on 98 models

Corrected

• Generic links replaced — 98 more dossiers linked to a company or university homepage. Open models from DeepSeek, LG (EXAONE), TII (Falcon), Baidu (ERNIE 4.5), Cohere (Aya, Command R), Mistral (Pixtral), Black Forest Labs (FLUX dev and schnell), AI21 (Jamba), BigCode, Tencent and others now link to their model cards; Bark, Moshi, OpenFold, Alpaca, ColBERT, Instant-NGP and the YOLO releases to their repositories; and OPT, Galactica, ULMFiT, RWKV-6, ERNIE 3.0 Titan, ModelScope T2V and the original latent diffusion model to their papers. All checked to load
• Open weights — Tencent’s Hy-MT2 translation models (Apache 2.0), Kyutai’s Helium-1 (CC BY-SA 4.0) and the second dots.llm1 row (MIT) publish their weights but were listed as proprietary; they now match their model cards
• Descriptions — Helium-1 was described as an audio model (it is a 2B text language model for edge devices), and the second dots.llm1 row called the model proprietary; both now follow their model cards

1,176 models · 329 vendors · 82 fields corrected on 82 models

Corrected

• Generic links replaced — 82 dossiers from Microsoft, Stability AI, Hugging Face, Amazon, xAI, Allen AI, Apple and DeepMind linked to a company homepage or index page. Phi models, Stable Diffusion XL and 3, SmolLM, BLOOM, Idefics, Tülu 3 and WildGuard now link to their model cards; StableLM, Grok-1, AlphaFold 2, MLX and Ferret to their repositories; ResNet, Phi-3, Gopher, Chinchilla, Gato, MuZero, ELMo and Apple’s foundation models to their papers; and Amazon Nova, Titan and Q Developer, the Grok releases and Microsoft’s Copilot products to their official pages. All checked to load

1,176 models · 329 vendors · 64 fields corrected on 42 models

Corrected

• Primary sources — 42 OpenAI dossiers linked only to OpenAI’s generic model index or homepage. API models now link to their own page in OpenAI’s developer docs, open and research releases (Whisper, CLIP, GPT-2, Jukebox, Shap-E, Point-E, gpt-oss) to their repository or model card, and GPT-3, InstructGPT, Codex and DALL·E to their papers
• Prices — 19 API models showed ChatGPT’s plan list or a placeholder instead of a price; they now show the per-token API price from OpenAI’s model page, marked “deprecated” where OpenAI marks them so (o1, o3, o4-mini, GPT-4, GPT-4 Turbo, GPT-3.5 Turbo, GPT-4.5 and others)
• Descriptions — o3, o4-mini and o3-mini, which had none, now have descriptions from OpenAI’s model pages

1,176 models · 329 vendors · 26 fields corrected on 26 models

Corrected

• EleutherAI’s X handle — five EleutherAI dossiers listed @EleutherAI; EleutherAI’s website and GitHub organization link @AiEleuther
• AceGPT — was listed under the United Arab Emirates, but it was released by CUHK-Shenzhen, the Shenzhen Research Institute of Big Data and KAUST; it now uses the value shared by other community releases
• Matching sibling rows — stock/status entries on Google, Anthropic, DeepSeek, TII and Recraft rows, and Alibaba’s CEO name, now match the other rows for the same company

1,176 models · 329 vendors · 26 fields corrected on 16 models

Corrected

• Open-weight models marked closed — TNG’s DeepSeek-R1T-Chimera, NVIDIA Cosmos 1.0, Spectra 1.1, International Moxin-7B and YaLM 100B publish their weights but were listed as proprietary; each now matches its model card or repository
• Closed models marked free — Yi-Large, Yi-Lightning, FLUX.1 [pro], FLUX1.1 [pro] Ultra, Luminous Supreme and Upstage Solar Pro were priced as free open-weight releases, though none has public weights; they now show their API pricing, or no price where none is published. Mubert, Opus Clip and Relevance AI, which are commercial apps, no longer claim open weights
• Smaller fixes — Gemini 2.5 Flash-Lite now shows its Gemini API free tier, and KAT-V1-200B, still in training, no longer shows Kuaishou’s Kling app subscription

1,176 models · 329 vendors · 127 fields corrected on 41 models

Corrected

• Open-weight models marked proprietary — 31 models whose weights are published on Hugging Face were listed as proprietary and cloud-only with an unknown free tier, so their dossiers said “Weights: API only”. Kimi K2, MiniMax-M1, Hunyuan-A13B, Reka Flash 3 and 3.1, dots.llm1, Dream 7B, Klear, Step3, Moonlight, Apertus, InternLM3, InternVL 2.5, Nemotron-4 and others are now marked open weights, free and locally deployable, each checked against its official model card and license
• Pricing — where those rows showed an unrelated or placeholder price (Kuaishou’s language models listed “Kling subscription”, the Kimi models a Kimi app subscription), it now reads “Free (open weights)”
• Free tier — 10 more open-weight models (including Qwen3-235B, Qwen2.5, three EXAONE models and DBRX) showed “No free tier” because the field was unknown; they now match the other open-weight rows

1,176 models · 329 vendors · 71 fields corrected on 71 models

Corrected

• Blank jurisdictions filled — 71 dossiers showed [UNKNOWN] for jurisdiction. Community releases now use the dataset’s “Open-source/community” value, BigCode models match their sibling row (France / USA), and companies and universities use the headquarters listed for them (for example Asana, Duolingo and Tesla in the USA, Bending Spoons in Italy, Lovable in Sweden, Monday.com and Wix in Israel, Naver in South Korea, Shopify in Canada, n8n in Germany)

1,176 models · 329 vendors · 59 fields corrected on 59 models

Corrected

• Generic links replaced — 59 dossiers linked to a generic page that doesn’t describe the model. Google research models (ViT, EfficientNet, ELECTRA, ALBERT, mT5, FLAN-T5, UL2, Switch Transformer and others) and the Gemma family pointed at the Gemini API site or the Gemini page, and Meta’s Code Llama, Seamless M4T, ImageBind, ESM, DINO and Detectron2 pointed at Meta’s generic research index. Each now links to its own paper, repository, model card or official page, all checked to load

1,176 models · 329 vendors · 60 fields corrected on 33 models

Corrected

• Descriptions added — 33 more dossiers that had no description now have one, written from the model’s paper or the vendor’s announcement: the Gemini Robotics family and Sec-Gemini, Cohere’s Command A and R7B, Mistral Saba, Microsoft Mu, Doubao-1.5-pro, Seed-Thinking-v1.5, Kimi k1.5, K2-Think, Pangu Ultra, Megatron-Turing NLG, wav2vec 2.0, YaLM 100B and AI21’s Jurassic models
• Primary sources — 27 of those dossiers linked only to a vendor homepage; they now link to the paper, announcement or repository (TimesFM-ICF’s link had pointed at the Gemini API docs)

1,176 models · 329 vendors · 96 fields corrected on 50 models

Corrected

• Descriptions added — 50 dossiers had no description (“What Makes It Special” showed [UNKNOWN]). Each now has one written from the model’s official model card or repository, including Kimi K2, MiniMax-M1, Hunyuan-A13B, Hermes 4, Mistral Small 3 and 3.1, QwQ-32B, OLMo 2 32B, Step3 and IBM’s Granite and Bamba models
• Primary sources — 43 of those dossiers linked only to a vendor homepage or docs index; they now link to the model’s own card or repository, all checked to load
• MAI-DS-R1 — Microsoft’s post-trained DeepSeek-R1 was listed as proprietary and not sold; its weights are published under the MIT license, so it is now marked open weights, free and locally deployable

1,176 models · 329 vendors · 252 fields corrected on 243 models

Corrected

• Placeholder removed — 234 dossiers showed “Private or parent-company status; verify from official source” as their stock/status. Each now shows a checked status: dated closes for listed companies and listed parents (Asana, Duolingo, HubSpot, Naver, Snap, Snowflake, Tesla and others), “Private” for 134 startups with no listing, and University, Non-profit or Research / open-source project for academic and community releases
• Listings and acquisitions — Bending Spoons (Remini) listed on Nasdaq in July 2026, Unitree on Shanghai’s STAR Market in August and Neosapience (Typecast) on KOSDAQ in September. Voyage AI now belongs to MongoDB, The Browser Company to Atlassian, PlayAI (Play.ht) to Meta, Tavily to Nebius and ELYZA to KDDI; Shutterstock stays listed after Getty Images ended their merger in July
• Consistency — undated “Public company” entries on Alibaba, Meta and Salesforce rows now use the same dated closes as their sibling rows
• CEOs — Groq’s CEO is now Simon Edwards (founder Jonathan Ross joined Nvidia); four rows that literally listed “Unknown” as CEO were cleared

1,176 models · 329 vendors · 20 fields corrected on 16 models

Corrected

• Copied descriptions — twelve models carried a sibling model’s description: Llama 3.3 and Llama 4 Maverick had the original Llama’s; Qwen3-Max, Qwen3-Max-Preview, Qwen3-Omni and Qwen3-Next had Qwen3-235B’s; DeepSeek-V3.1 and V3.1-Terminus had DeepSeek-V3’s; Grok 4 Fast had Grok 4’s; two Gemini 2.5 previews had the March release’s; and GPT-NeoX had GPT-Neo’s. Each now describes its own model, from the vendor’s announcement or model card
• Wrong claims — Grok 4’s description gave a 2M-token context window, which is Grok 4 Fast’s (xAI lists 256K for the Grok 4 API); gpt-oss-120b and gpt-oss-20b were described as OpenAI’s “first open-source model,” though GPT-2’s weights, Whisper and CLIP came earlier
• Qwen3-Max — Qwen3-Max and Qwen3-Max-Preview were marked open source; Alibaba never published their weights, so both are now proprietary and cloud-only
• Gemini Robotics-ER 1.6 — its stock field showed a bare “333”; it now reads $344 (GOOGL, Sep 25, 2026), matching the other Google DeepMind rows
• Homepage list — the static model list on the homepage now shows DeepSeek-V4-Flash’s corrected description from September 29

1,176 models · 329 vendors · 254 source records corrected on 171 models

Corrected

• Release sources — 135 release-source links were dead or had been copied from an alphabetically neighboring vendor (Tabnine pointed at Suno’s blog, Tesla Optimus at Tencent’s Hugging Face page, VGG16 at Udio). They now point at the model’s paper, repository or official announcement, or were cleared where no official page exists
• Pricing and docs — 71 pricing-source and 48 docs links pointed at another vendor’s pages (EleutherAI models at OpenAI’s docs, community Mistral fine-tunes at Mistral’s pricing); each was replaced with the model’s own repository or cleared
• Where these live — these source columns back the dataset’s verification and aren’t shown on dossier pages

1,176 models · 329 vendors · 52 fields corrected on 51 models

Corrected

• Wrong-organization links — 51 primary-source links pointed at another organization’s site or an unrelated product. Meta research models (Segment Anything, MusicGen, NLLB-200, RoBERTa) and other labs’ Llama derivatives (NVIDIA Nemotron, Ai2’s Tülu 3, TinyLlama) linked to Meta’s Llama page; AlphaGo, AlphaZero, Med-PaLM and PaLM-E to the Gemini page; EleutherAI’s GPT-J, GPT-Neo and GPT-NeoX to OpenAI’s docs; GLM-5.1 to GLM-5’s model card. Each now links to its own repository, model card or official page, all checked to load
• Cerebras Llama 3.3 70B — now marked retired: Cerebras deprecated it on February 16, 2026

1,176 models · 329 vendors · 53 fields corrected on 5 models

Corrected

• Samsung Galaxy AI — carried Meta’s staff, ticker, X handle and US jurisdiction. Now: Samsung’s leadership, South Korea, ₩276,000 (005930.KS, Oct 1, 2026), @Samsung, free basic features and the Galaxy AI page
• llama.cpp — listed Meta’s executives, stock and X handle. It’s an independent open-source project created by Georgi Gerganov (ggml.ai, acquired by Hugging Face in 2026); now @ggml_org and its GitHub repository
• GPT4All — showed ChatGPT’s pricing; Nomic’s app is free and open source
• Inception Mercury and Mercury Coder Small — carried data from G42’s unrelated Inception (UAE). Mercury is made by Inception of Palo Alto (CEO Stefano Ermon); jurisdiction, CEO, X handle, pricing, descriptions and links corrected

1,176 models · 329 vendors · 41 primary-source links repaired

Corrected

• Link check — all 547 unique primary-source links were tested; 41 rows pointed at pages that no longer load (404s, dead domains, SSL failures)
• Repaired — 38 now point at the working official page (for example the Gemini 1.0 launch post, the DBRX announcement, Segment Anything’s repository and Stability AI’s site); 3 dead links with no working replacement (GeoGPT, Play.ht, Tome) were removed

1,176 models · 329 vendors · 115 X handles corrected

Corrected

• X handles — now match the account each company links from its own website: Meta @AIatMeta (70 models), Ai2 @allen_ai, Moonshot/Kimi @Kimi_Moonshot, Black Forest Labs @bfl_ai, Kling @Kling_ai, Ideogram @ideogram_ai, Suno @suno and ElevenLabs @elevenlabs

1,176 models · 329 vendors · 3 release dates added

Added

• Release dates — Magistral Medium (June 10, 2025), Nemotron Nano 8B (March 18, 2025) and TNG DeepSeek R1T Chimera (April 27, 2025), each from the official release page or model card

1,176 models · 329 vendors · 2 fields corrected

Corrected

• DeepSeek V4 Flash — the description said the July 31, 2026 release was API-only with no weights; DeepSeek published MIT-licensed weights on Hugging Face with that release
• Devstral Medium — release date added (July 10, 2025)

1,176 models · 329 vendors · 315 fields corrected on 314 models

Corrected

• Deployment — 313 open-weight models (gpt-oss, Qwen3, Gemma 3, Llama 3.x, DeepSeek distills and others) were listed as “Cloud” only; they now read “Cloud + Local (open weights)”
• Devstral Medium — not open weights: Mistral offers it through its API (with on-premise deployment for enterprises); only Devstral Small 1.1 is Apache 2.0

1,176 models · 329 vendors · 71 pricing fields made specific

Corrected

• Meta (56 rows) — “Mostly free/open models…” replaced with each model’s access: open weights (Llama models note the Llama license), free in the Meta AI app, or not sold for research-only models
• Mistral (15 rows) — “Le Chat / API pricing varies” replaced with open weights, Le Chat plus API, or API-only pricing

1,176 models · 329 vendors · 34 fields corrected on 14 models

Corrected

• Now marked open weights — GPT-1, GPT-2 and GPT-2 1.5B, XLNet, Xception, Inception v3, OPT, Galactica, ERNIE 4.5 21B-A3B Thinking and GLM-4.5 (all have publicly released weights); Chameleon is marked partial (research-only license)
• Deployment fixed — Mistral Small 3 and 3.1 (Apache 2.0) and AudioLDM 2 were listed as “cloud only” despite open weights

1,176 models · 329 vendors · 113 pricing fields corrected

Corrected

• Bulk-filled app pricing removed — 113 research, open-weight and API-only models from OpenAI, Google, Google DeepMind, DeepMind, xAI and others carried ChatGPT, Google AI Pro or X Premium subscription prices
• Now — 63 read “Free (open weights)”, 33 “Not sold”, 9 “Retired” (GPT-3, InstructGPT, Codex, PaLM), 4 show API pricing and 4 are blank pending a sourced price

1,176 models · 329 vendors · 49 pricing fields updated

Corrected

• Copilot — Copilot Pro was discontinued in October 2025; consumer Copilot now shows free Copilot and Microsoft 365 Premium at $19.99/month, and Microsoft 365 Copilot and Security Copilot show their business pricing
• Microsoft research and open models — 40 rows (Phi, ResNet, DeBERTa, Swin and others) had consumer Copilot pricing bulk-filled in; they now read “Free (open weights)” or “Not sold (research model)”

1,176 models · 329 vendors · 7 stock fields updated

Corrected

• Yandex — listed on the Moscow Exchange (YDEX) since July 2024; the old “Delisted” described the former Dutch parent’s suspended Nasdaq shares
• Xiaohongshu (RedNote) — the $17B figure is now dated to its 2024 round, alongside secondary trades of up to about $50B in late 2025

1,176 models · 329 vendors · 59 stock fields updated

Corrected

• Mistral AI — valued at more than €21B after its September 2026 Series D; the old “$6B (2025)” was really the June 2024 valuation
• xAI — the SpaceX listing note now carries a dated share price (SPCX close, September 28, 2026)
• Kuaishou — refreshed with the September 29, 2026 close; the old HKD 55 dated from April 2026

1,176 models · 329 vendors · 14 fields updated on 8 models

Corrected

• Zhipu AI — its Qingyan app generates images and video; GLM-4, GLM-5, 5.1, 5.2, 5.3 and 5.3-Flash now show Image and Video generation
• Pika — the $470M valuation is now dated to its June 2024 Series B

1,176 models · 329 vendors · 12 capability flags corrected

Corrected

• MiniMax — MiniMax’s own Agent app produces images, video and audio; MiniMax-M2, M2.5, M2.7 and M3 now show Image, Video and Audio generation

1,176 models · 329 vendors · 6 capability flags corrected

Corrected

• Voice Chat — Kimi’s app offers voice calls; Kimi Chat, K2 Thinking, K2.5, K2.6 and K3 now show Voice Chat
• Image Gen — Tencent’s Yuanbao app (Hunyuan Chat) generates images

1,176 models · 329 vendors · 4 stock fields updated

Corrected

• Together AI — valued at $8.3B after its July 2026 Series C; the old $3.3B dated from February 2025
• SambaNova — valued at about $11B after its July 2026 round; the old $5.1B dated from 2021

1,176 models · 329 vendors · 20 fields updated on 12 models

Corrected

• Runway — valued at $5.3B after its February 2026 Series E; the old “$4B+” figure is gone and all eight Runway rows now carry the same dated valuation
• LG (EXAONE) — LG Electronics CEO is Lyu Jae-cheol, who succeeded William Cho on December 1, 2025

1,176 models · 329 vendors · 10 stock fields updated

Corrected

• Suno — valued at $5.4B after its June 2026 Series D; the old $500M dated from May 2024
• 01.AI — the $1B+ valuation is dated November 2023; now also notes its planned 2027 Hong Kong IPO

1,176 models · 329 vendors · 24 stock fields updated

Corrected

• Hugging Face — last priced at $4.5B (August 2023); now noted as exploring a sale at $13B+ (August 2026)
• Moonshot AI — valued at $35B after its July 2026 round; the old “$3B+” dated from 2024
• Databricks — valued at $190B after its August 2026 round; the old $62B dated from December 2024

1,176 models · 329 vendors · 24 fields updated on 22 models

Corrected

• Cohere and Aleph Alpha — signed a definitive merger agreement on September 16, 2026 (combined value about $20B, pending regulators)
• ByteDance — valued at $330B+ in its August 2025 employee buyback; the old $268B figure dated from December 2023
• Anysphere (Cursor) — last priced at $29.3B (November 2025), with a $60B SpaceX acquisition agreement reported in 2026; the old “$10B+” was stale

1,176 models · 329 vendors · 421 stock fields updated

Corrected

• Public companies — every stock price now carries its ticker and the date of the closing price it reflects (latest closes as of September 25–28, 2026); undated figures are gone
• Zhipu AI and MiniMax — now listed as public companies; both began trading on the Hong Kong Stock Exchange in January 2026
• Windsurf — acquired by Cognition in July 2025; the old “OpenAI acquisition announced” status is removed

1,176 models · 329 vendors · 46 cells corrected on 21 models

Corrected

• Nous Research — CEO is Jeffrey Quesnelle; Karan Malhotra is Head of Behavior
• Prime Intellect — CEO is co-founder Vincent Weisser
• AI21 Labs — led by co-CEOs Ori Goshen and Yoav Shoham, with Amnon Shashua as chairman
• Image Gen — Perplexity and Baidu’s ERNIE Bot generate images in their apps
• Claude Mythos 5 and 5.1 — PDF and spreadsheet marks removed; Mythos-tier models are not offered in Anthropic’s consumer apps

1,176 models · 329 vendors · 16 cells corrected on 4 models

Corrected

• Windsurf — owned by Cognition (CEO Scott Wu) since July 2025; former CEO Varun Mohan joined Google
• GitHub Copilot — GitHub has had no CEO since Thomas Dohmke left at the end of 2025; it reports into Microsoft CoreAI
• Aleph Alpha — co-CEOs Reto Spörri and Ilhan Scheer; founder Jonas Andrulis stepped down in 2025

1,176 models · 329 vendors · 69 cells corrected on 25 models

Corrected

• Apple — CEO is now John Ternus (since September 1, 2026); Tim Cook is executive chairman
• Yandex — CEO is Artyom Savinovsky; founder Arkady Volozh left with the 2024 sale of the Russian business
• TII (Falcon) — CEO is Dr. Najwa Aaraj; the previous entry named the head of TII’s parent council
• Release dates — Gemini Robotics 1.5 and Robotics-ER 1.5, Hermes 4, Apertus, Step 3, Windsurf SWE-1 and Cerebras Llama 3.3 70B now carry their verified release dates

1,176 models · 329 vendors · 254 capability cells corrected on 93 models

Corrected

• App Available — set to No on 72 research, open-weight and API-only models that have no consumer app (for example BERT, GPT-2, T5, PaLM, Gemma, Phi and the Gemini Robotics family)
• PDF, Spreadsheet and Voice flags — removed where they had been bulk-filled onto models that cannot produce them (44 PDF, 43 spreadsheet, 64 voice)
• Qwen Chat — Qwen models released in 2025 or later now show image, video and voice generation
• Copilot and Doubao — Copilot shows image creation and voice; Doubao shows image, video, music and voice

1,176 models · 329 vendors · 183 capability flags corrected on 89 models · 90 safety placeholders removed

Corrected — one definition for the capability columns

• Definition — a capability is marked Yes when a user can get that output for the model in the vendor’s own app, including the app’s built-in tools; third-party add-ons do not count
• Image Gen — 82 models now marked Yes (ChatGPT, Gemini, Grok, Meta AI and Le Chat all generate images in-app)
• Video Gen — 33 models now marked Yes (Gemini, Grok and Meta AI generate video in-app)
• PDF Gen and Spreadsheet Gen — 18 and 17 models now marked Yes (ChatGPT and Claude file creation)
• Voice Chat — 33 models now marked Yes (Claude, ChatGPT, Gemini Live, Grok and Meta AI voice modes)
• Known Safety Issues — “None publicly documented” and “None widely reported” placeholders removed; blank now means no sourced incident on file

1,140 → 1,176 models · 329 vendors · 36 new models · 34 source-backed corrections

Added — 36 releases, primary-source verified

• Alibaba — Qwen3-VL, Qwen-Image
• Allen AI — Olmo 3
• Amazon — Amazon Nova 2 Lite, Amazon Nova 2 Omni, Amazon Nova 2 Sonic
• Anthropic — Claude Opus 5.5, Claude Mythos Preview
• Black Forest Labs — FLUX.1 Kontext, FLUX.2
• DeepSeek — DeepSeek-V3.2-Speciale, DeepSeek-V3.2-Exp
• Google DeepMind — EmbeddingGemma, TranslateGemma, Gemini 2.5 Flash-Lite
• IBM — Granite 4.0
• Kuaishou — Kling 2.5 Turbo
• Meta AI — V-JEPA 2
• Microsoft — MAI-Voice-1
• Midjourney — Midjourney V1 Video, Midjourney V8.1
• MiniMax — MiniMax-M2.1
• Mistral AI — Voxtral, Ministral 3, Mistral OCR 3
• NVIDIA — Nemotron 3 Nano, Nemotron 3 Super
• OpenAI — GPT-6 Sol, GPT-6 Luna, GPT-5 mini, GPT-5 nano, GPT-5.4 mini, GPT-5.4 nano
• xAI — Grok 4.7, Grok Code Fast 1
• Zhipu AI — GLM-4.7

Corrected — 34 rows, each tied to a vendor or primary source

• Release dates — Gemma 2, CodeGemma, ShieldGemma, T5Gemma, Gemma 3 270M, VaultGemma; ERNIE X1.1, ERNIE 4.5 21B-A3B Thinking, ERNIE 5; EXAONE 4.0, EXAONE Deep, EXAONE 4.5; Stable Diffusion 3.5; Runway Gen-4; Kling 2.0; Suno v4; MAI-1-preview; SAM 3; GLM-4.5; Codestral Mamba; Devstral Medium; Midjourney v5.2; FLUX.1 [dev]/[pro]
• Licensing — Mistral Large 3 is Apache 2.0 open weights
• Sources & cleanup — wrong-product primary links fixed on Gemma rows; missing DeepSeek-V3.2 and pricing sources added; OpenAI fields removed from GPT-J-6B, whisper.cpp and GPT4All
• Ratings — the 33 September 14 additions recomputed with the published v2.0 formula (Capability 50% · Currency 20% · Access 15% · Significance 15%); 12 ratings changed

1,102 → 1,140 models · 324 → 329 vendors · 38 new models · 5 new vendors · dataset-wide field refresh (data completeness, CEO politics, verification dates)

Added — 38 releases, internet-verified

• Alibaba — Qwen3.8-Max-0902, Qwen3.8-Flash-Next, Qwen3.8-27B
• Ant Group — Ling-3.0-Flash-Fin, Ling-3.0-Flash-Sante
• Anthropic — Claude Fable 5.1, Claude Mythos 5.1
• DeepSeek — DeepSeek-V4.1-Flash, DeepSeek-V4-Flash-Vision-Exp
• ElevenLabs — Eleven Music
• Google DeepMind — Gemini 3.8 Flash, Gemini 3.8 Flash Cyber, Gemini 3.7 Flash, Nano Banana 2
• IFM — K2 Horizon 375B-A23B, K2 Horizon 36B-A4B, K2 Horizon 32B, K2 Horizon 7B, K2 Horizon 3.7B, K2 Horizon 0.9B
• LG — EXAONE 4.5
• Lightricks — LTX-Video
• Meta AI — Muse Spark 1.3, Muse Glimmer
• OpenAI — GPT Image 2.5 Flare, GPT Image 2.5 Sunburst, GPT-6 Astra
• Peking University / Kuaishou — Pyramid Flow
• Recraft — Recraft V4
• Rhymes AI — Allegro
• Skywork AI — SkyReels V1
• StepFun — Step-Video-T2V
• Suno AI — Suno v5.5
• Tencent — Hy-MT2-30B-A3B, Hy-MT2-1.8B
• xAI — Grok 4.6
• Zhipu AI — GLM-5.3-Flash, GLM-5.3

1,037 → 1,102 models · 313 → 324 vendors · 65 new models (Jun 10 – Aug 9), 101 field updates, CEO-politics column overhauled, 8 rescores

Added — 65 releases, internet-verified

• Anthropic — Claude Sonnet 5 (Jun 30) and Claude Opus 5 (Jul 24)
• OpenAI — GPT-5.6 family: Sol, Terra, Luna (Jul 9)
• xAI — Grok 4.5 · Moonshot AI — Kimi K3 (2.8T, largest open-weights release to date) · Zhipu — GLM-5.2 · Alibaba — Qwen3.8-Max
• Inkling — Inkling and Inkling-Small (new vendor) · Meta — Muse family · Microsoft — 8 MAI models
• Video backfills: Runway Gen-4.5, Luma Ray3, Kling 3.0 · 11 new vendors onboarded with full profiles

Corporate events

• xAI → SpaceX merger reflected across all xAI rows (Nasdaq: SPCX, @SpaceXAI)
• Google DeepMind — Demis Hassabis stepped down Aug 5; Koray Kavukcuoglu now runs day-to-day (81 rows updated)
• Anthropic — confidential IPO filing (Jun 1) · Meta — MSL leadership under Wang / Friedman

Corrected & rescored

• CEO-politics column standardized on 226+ rows — documented activity only; PRC-based executives labeled uniformly
• Sora / Sora 2 retired (Currency → 2) · 7 superseded flagships rescored (Currency → 6)
• Model dossiers now render the AIDb Rating on its native 10-point scale (previously mislabeled “/5”)

1,037 models · 313 vendors · 10 cells corrected, dossiers upgraded

• Stability AI — stale CEO cell updated to Prem Akkaraju
• 9 stock-status placeholders filled by strict sibling inheritance (BAAI ×3, Aleph Alpha, Canva, Coqui, Sber, Upstage, Xiaohongshu); 234 remain pending manual verification
• Dossiers now show the AIDb Rating with full score basis and a Primary source link with per-model Verified date

1,041 → 1,037 models · 316 → 313 vendors · full-database integrity audit: zero duplicates, zero encoding defects, arithmetically verified scores

Removed — duplicate rows retired

• ERNIE 3.0 Titan (Baidu) — duplicate row retired; rated entry kept, description restored
• Granite 13B (IBM) — duplicate row retired; rated entry kept
• DALL-E 3 (OpenAI) — duplicate row retired; rated entry kept
• MobileNetV3 (Google) — byte-identical twin retired

Corrected

• Kandinsky 3 (Sber) — jurisdiction corrected from China to Russia
• GPT-4.1 — description rewritten (previously carried GPT-4's 2023 origin blurb)
• mmBERT (JHU) — description rewritten (previously carried BERT's 2018 blurb)
• Llama 3.1 Nemotron Ultra 253B & Llama 3.3 Nemotron Super 49B (NVIDIA) — descriptions rewritten (previously carried Meta's Llama origin blurb)
• Nucleotide Transformer — description rewritten (previously carried the "Attention Is All You Need" blurb)
• Vendor normalization: Minimax → MiniMax, Nvidia → NVIDIA, TNG Technology → TNG Technology Consulting

Site

• Machine-readable layer shipped: robots.txt, llms.txt, sitemap.xml (1,038 URLs), and the full dataset at models.json
• Per-model structured data (schema.org) now on every dossier page, including the AIDb rating as a formal review
• Compare view fixed — App and Data Completeness rows now render on desktop and mobile; Dictation row live
• Rolodex completeness bars now reflect curated Data Completeness values
• All site references updated to July 2026

1,002 → 1,041 models (+49 / −10) · 316 vendors · every published row re-verified with per-model dates and primary source URLs

Highlights among the 49 additions

• Claude Fable 5 & Claude Mythos 5 (Anthropic) — first Claude 5 family entries, plus Claude Opus 4.8
• GPT-5.5, GPT-5.5 Pro & GPT-5.3-Codex (OpenAI)
• Gemini 3.5 Flash, Gemini 3 Deep Think, Gemma 4, Veo 3.1 & the Nano Banana line (Google DeepMind)
• Grok 4.3, 4.20 & 4.1 (xAI) · DeepSeek-V4 Pro/Flash · Kimi K2.6/K2.5 (Moonshot) · GLM-5.1/5 (Zhipu) · Qwen3.7-Max line (Alibaba) · MiniMax-M3
• New vendors: Krea AI (Zing) and Loudly

Removed

• 7 flagged rows: Claude 4.1 umbrella (dup of Opus 4.1), Llama 4 Behemoth (unreleased), Mistral Small 24B (dup), Kimi K1.5 & K2 vendor dups, gpt-oss & o3/o4-mini umbrella rows
• 3 junk rows: a corrupted DeepSeek R1-0529 duplicate, a broken pixart- id, and an "Unknown"-vendor ModernBERT twin

Rescore — AIDb v2.0

425 models rated 1–10, each with a published, verifiable score basis:

Data

• New capability flags: PDF Gen, Spreadsheet Gen, Dictation
• Per-row Last Verified dates (2026-05-12 / 2026-06-11 passes) and primary source URLs on 98.7% of rows