Data quality

What this dataset can prove, and what it cannot · 280 records · updated 2026-10-15

Every figure on this page is read back out of the dataset itself. Where a record cannot support a claim, it says so here rather than quietly rounding up.

Verification

A record is verified when every value it asserts was found in a primary source — one published by the organisation that made the model. A null asserts nothing, so an undisclosed figure does not block verification.

Sources

570 citations across 280 records. 545 are primary. 528 carry a dated archive snapshot, so they still prove the claim after the live page changes.

Every source URL is re-checked weekly in CI. 2 records cite no primary source at all and cannot be verified until the lab's own announcement is found: Llama 5, Muse Spark 1.2.

Specification coverage

A missing value is only a gap if someone could have recorded it. Not disclosed means we read the primary sources and the lab does not publish it — the record is complete. Not researched means nobody has checked yet. Most proprietary labs never publish parameter counts, so those nulls are correct answers, not holes.

Pricing and benchmarks are thin, and deliberately so. A price is true for a quarter and a leaderboard changes weekly, while every figure here must be traced to a dated primary source before it can be published — the wrong shape for live data. OpenRouter and Artificial Analysis maintain those properly, and each model page links out to them.

Modalities are new to schema 1.6. The earlier schema recorded that a model was multimodal but never which modalities, so these are being researched from primary sources rather than back-filled with assumptions.

Fact-level evidence

A record-level status says the record was checked. This says which source states which number, so a figure can be traced without taking the badge on trust. Three fields carry it so far — doing every field at once is what makes this kind of work never ship.

No record currently has sources that disagree with each other. Where that happens, both values are published rather than one being chosen silently.

Pricing and benchmarks

Both were nearly left out of this project, on the grounds that a price changes silently and a citation to a live page rots. The answer was to cite dated snapshots instead: a pricing entry whose source has no archive is rejected by the build, so every price here is evidenced as of a specific capture.

Most labs do not state a price in the announcement itself, and this dataset does not chase the current price on a live pricing page — that would be a number without a date, which is the thing it is trying not to publish.

Verification by lab

Share of each lab's tracked releases that are fully verified. Labs that publish detailed model cards verify faster; that is a fact about them, not about their models.

What is not covered

This dataset tracks 280 releases from 27 labs. It is not a complete census and does not imply one. Notable absences include Ai2 (OLMo), TII (Falcon), Baidu, ByteDance, Reka and Stability, along with most fine-tunes, quantisations and community derivatives, which are out of scope by design. Naming the gaps is more useful than a number that implies there are none.

Records not yet verified

95 of 280. Each says which fact is unproven.

Nothing matches those filters.

1. EmbeddingGemma 2 Cited to the lab's own announcement. Not verified because the release date and the parameter count could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. The primary source states the model is text-only. Capabilities evidenced from the primary sources: video: “6, 2026 Google AI Edge Team Android ML Team Share Facebook Twitter LinkedIn Mail Today, Google DeepMind launched EmbeddingGemma 2 , a best-in-class for its size, open-weight, multi”.
2. Claude Haiku 5.5 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
3. Muse Spark Open Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
4. Claude Fable 5.1 Cited to official documentation and the lab's own announcement. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
5. Claude Mythos 5.1 Cited to official documentation and the lab's own announcement. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
6. Gemini 3.8 Flash Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`. Downgraded: context_window comes from the lab's documentation, but attribute-facts could not read an archived snapshot that states it, so the figure is not yet traced to a quoted sentence. reconcile-status upgrades it once attribute-facts succeeds.
7. Gemini 3.8 Flash Cyber Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`. Downgraded: context_window comes from the lab's documentation, but attribute-facts could not read an archived snapshot that states it, so the figure is not yet traced to a quoted sentence. reconcile-status upgrades it once attribute-facts succeeds.
8. GPT-6 Astra Cited to the lab's own announcement and official documentation. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
9. GPT-Rosalind Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. The primary source states the model is text-only.
10. TS Jev Cited to the lab's own announcement and official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Capabilities evidenced from the primary sources: function_calling: “Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”.
11. Qwen3.8-Omni-Flash Cited to the lab's own announcement. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2026-09-17, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-09-18 would settle it. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
12. Qwen3.8-LiveTranslate Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
13. Qwen-Image-2.1 Cited to the lab's own announcement and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities from the model card's own pipeline_tag "text-to-image" on Qwen/Qwen-Image-2.1.
14. Claude Opus 5.5 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
15. GPT-6 Sol Cited to official documentation and the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. The primary source states the model is text-only.
16. GPT-6 Luna Cited to official documentation and the lab's own announcement. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
17. Gemini 3.8 Flash TTS Cited to official documentation. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2026-09-21, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-09-22 would settle it.
18. Gemini 3.8 Flash-Lite TTS Cited to official documentation. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2026-09-21, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-09-22 would settle it.
19. MiniMax M3.1-Flash-Preview Cited to the lab's own announcement and official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
20. Claude Sonnet 5.5 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
21. GPT-6.1 Sol Cited to official documentation. Not verified because the release date and the context window could not be traced: every archived snapshot of those sources predates the release — the newest is 2026-09-28, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-09-29 would settle it.
22. Qwen3.8-Max Re-sourced to the lab's own announcement, replacing a news citation. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present. The promised weights were published on 2026-08-08 as Qwen3.8-2.4T-A95B and are tracked as their own record.
23. Qwen3.8-Flash-Next Cited to the lab's own announcement and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities from the model card's own pipeline_tag "image-text-to-text" on Qwen/Qwen3.8-Flash-Next.
24. Muse Spark 1.2 No primary source. The release is recorded from news reporting only, so it cannot be verified until the lab's own announcement is cited.
25. Qwen3.8-27B Cited to the official model card. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities from the model card's own pipeline_tag "image-text-to-text" on Qwen/Qwen3.8-27B.
26. Qwen3.8-2.4T-A95B Cited to the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Licence from the model card's own license_name field on Qwen/Qwen3.8-2.4T-A95B.
27. MAI-Thinking-1 Cited to the lab's own announcement. the release date could not be traced yet: this release is 1 day old and archive.org has not captured the announcement since it went up. That is a wait, not a gap — re-running the archive pass will settle it.
28. OR Ornith-1.5-397B Cited to the lab's own announcement and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
29. DeepSeek-V4-Flash-Vision-Exp Cited to official documentation and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities from the model card's own pipeline_tag "image-text-to-text" on deepseek-ai/DeepSeek-V4-Flash-Vision-Exp.
30. Gemini 3.5 Transcribe Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
31. Gemini 3.5 Transcribe Live Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
32. Seedream 5.0 Pro Cited to the lab's own announcement and official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
33. GPT-5.6 Sol Split from the former "gpt-5-6" record, which was named for the family while holding one variant's figures. Evidence: "GPT-5.6 Sol is the frontier model in the GPT-5.6 series." Modalities taken from the documentation's own declaration: "Input Text, image Output Text". Capabilities evidenced from the primary sources: reasoning: “Reasoning token support”.
34. GPT-5.6 Luna Split from the former "gpt-5-6" record, which was named for the family while holding one variant's figures. Evidence: "GPT-5.6 Luna Default — its own model page in OpenAI's API documentation, with its own specifications." Modalities taken from the documentation's own declaration: "Input Text, image Output Text". Capabilities evidenced from the primary sources: reasoning: “Reasoning token support”.
35. GPT-5.6 Terra Split from the former "gpt-5-6" record, which was named for the family while holding one variant's figures. Evidence: "GPT-5.6 Terra Model — its own model page in OpenAI's API documentation, with its own specifications." Modalities taken from the documentation's own declaration: "Input Text, image Output Text". Capabilities evidenced from the primary sources: reasoning: “Reasoning token support”.
36. Qwen-Image-3.0 Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
37. Seed Audio 1.0 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
38. Gemini 3.5 Flash-Lite Cited to official documentation. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
39. MiniMax H3 Cited to the official model card and the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Licence from the model card's own license_name field on MiniMaxAI/MiniMax-H3. Capabilities evidenced from the primary sources: video_generation: “case-T2VA Type: Text-to-video Duration: 10 seconds Aspect ratio: 16:9 stage request result H3-Context-IR View script { "task": { "id": "<task_id>", "model": "MiniMa”.
40. Gemini Robotics-ER 2 Cited to official documentation. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2026-07-29, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-07-30 would settle it.
41. Seedance 2.5 Cited to the lab's own announcement and official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
42. Claude Mythos 5 Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Capabilities evidenced from the primary sources: multilingual: “All current Claude models support text and image input, text output, multilingual capabilities, and vision.” | vision: “All current Claude models support text and image input, text output, multilingual capabilities, and vision.”.
43. GLM-5.2 Cited to the lab's own announcement and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Zhipu's own release notes date GLM-5.2 to 2026-06-16 where this record publishes 2026-06-13; the disagreement is recorded rather than resolved.
44. Seed2.1 Cited to the lab's own announcement and official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
45. Gemini Omni Flash Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
46. Nano Banana 2 Lite Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
47. Gemini 3.1 Flash-Lite Cited to official documentation. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
48. Qwen3.7-Max Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
49. Qwen3.7-Plus Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
50. Mistral Medium 3.5 Cited to the lab's own announcement. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2026-05-21, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-05-22 would settle it.
51. Nano Banana Pro Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
52. Grok Build 0.1 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`. Downgraded: context_window comes from the lab's documentation, but attribute-facts could not read an archived snapshot that states it, so the figure is not yet traced to a quoted sentence. reconcile-status upgrades it once attribute-facts succeeds.
53. GLM-5V-Turbo Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
54. Qwen3.6-Plus Cited to the lab's own announcement. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
55. Llama 5 No primary source. The release is recorded from news reporting only, so it cannot be verified until the lab's own announcement is cited.
56. Muse Spark Cited to the lab's own announcement. Not verified because the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
57. Seed3D 2.0 Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
58. Grok 4.20 Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
59. Grok 4.20 Multi-agent Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
60. GLM-5-Turbo Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
61. Mistral Small 4 Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
62. Lyria 3 Cited to official documentation. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2025-06-05, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2026-03-25 would settle it.
63. Gemini 3.1 Flash Live Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
64. Veo 3 Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
65. Qwen3-Coder-Next Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
66. GLM-OCR Cited to official documentation and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Licence from the model card's own license field on zai-org/GLM-OCR. Modalities from the model card's own pipeline_tag "image-to-text" on zai-org/GLM-OCR.
67. GPT-5.3-Codex Cited to official documentation. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities read from OpenAI's own model documentation, which states: “Modalities Text Input and output Image Input only Audio Not supported Video Not supported.” The snapshot cited was moved from 2026-02-16 to 2026-02-26: the earlier capture caught only the page shell before its content rendered, so it evidenced nothing it was cited for. Capabilities evidenced from the primary sources: vision: “Modalities Text Input and output Image Input only Audio Not supported Video Not supported”.
68. GLM-5 Cited to official documentation and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. The primary source states the model is text-only. Licence from the model card's own license field on zai-org/GLM-5.
69. Seedance 2.0 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
70. Seedream 5.0 Lite Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
71. Qwen3.5-397B-A17B Cited to the lab's own announcement and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Licence from the model card's own license field on Qwen/Qwen3.5-397B-A17B.
72. Seed2.0 Newly added. Sources are not archived and no value has been traced to one yet — run `npm run enrich`.
73. GLM-Image Cited to official documentation and the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Licence from the model card's own license field on zai-org/GLM-Image. Modalities from the model card's own pipeline_tag "text-to-image" on zai-org/GLM-Image.
74. MiniMax Speech 2.8 Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
75. Qwen3-Max-Thinking Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
76. Mistral Large 3 Split from the former "mistral-3" record, which was named for the family while holding one variant's figures. Evidence: "Mistral Large 3 – our most capable model to date – a sparse mixture-of-experts trained with 41B active and 675B total parameters." Capabilities evidenced from the primary sources: multilingual: “Introducing Mistral 3 | Mistral AI Products Solutions Research Resources Pricing Company Try AI Studio Talk to sales Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mis”.
77. Ministral 3 3B Split from the former "mistral-3" record, which was named for the family while holding one variant's figures. Evidence: "Ministral 3 series, available in three model sizes: 3B, 8B, and 14B parameters." Capabilities evidenced from the primary sources: multilingual: “Introducing Mistral 3 | Mistral AI Products Solutions Research Resources Pricing Company Try AI Studio Talk to sales Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mis”.
78. Ministral 3 8B Split from the former "mistral-3" record, which was named for the family while holding one variant's figures. Evidence: "Ministral 3 series, available in three model sizes: 3B, 8B, and 14B parameters." Capabilities evidenced from the primary sources: multilingual: “Introducing Mistral 3 | Mistral AI Products Solutions Research Resources Pricing Company Try AI Studio Talk to sales Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mis”.
79. Ministral 3 14B Split from the former "mistral-3" record, which was named for the family while holding one variant's figures. Evidence: "Ministral 3 series, available in three model sizes: 3B, 8B, and 14B parameters." Capabilities evidenced from the primary sources: multilingual: “Introducing Mistral 3 | Mistral AI Products Solutions Research Resources Pricing Company Try AI Studio Talk to sales Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mistral 3 Mis”.
80. Nemotron 3 Nano 30B-A3B Cited to the technical paper. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2025-12-17, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2025-12-23 would settle it. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
81. GPT-5.1 Cited to official documentation. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities taken from the documentation's own declaration: "Input Text, image Output Text". Capabilities evidenced from the primary sources: reasoning: “Reasoning token support”.
82. Claude Haiku 4.5 Cited to the lab's own announcement and official documentation. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
83. Sora 2 Cited to the lab's own announcement. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
84. Imagen 4 Cited to official documentation. Not verified because the release date could not be traced: every archived snapshot of those sources predates the release — the newest is 2025-06-05, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2025-08-14 would settle it.
85. GLM-4.5 Cited to the lab's own announcement and the official model card. Not verified because the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
86. DeepSeek-R1 Cited to the technical paper and the official model card. Not verified because the context window and the parameter count could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present. Capabilities evidenced from the primary sources: reasoning: “View a PDF of the paper titled DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning, by DeepSeek-AI and 199 other authors View PDF HTML (exp”.
87. Qwen2-72B Cited to the lab's own announcement and the official model card. Not verified because the parameter count could not be traced: every archived snapshot of those sources predates the release — the newest is 2024-06-06, captured before the page announced it — so the archived copy does not contain the figure. A capture from on or after 2024-06-07 would settle it.
88. GLM-4 Split from the former "glm-4" record, which was named for the family while holding one variant's figures. Evidence: "This report primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B." Capabilities evidenced from the primary sources: multilingual: “7k GLM-4 series: Open Multilingual Multimodal Chat LMs | 开源多语言多模态对话模型 License Apache-2.”.
89. GLM-4-Air Split from the former "glm-4" record, which was named for the family while holding one variant's figures. Evidence: "This report primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B." Capabilities evidenced from the primary sources: multilingual: “7k GLM-4 series: Open Multilingual Multimodal Chat LMs | 开源多语言多模态对话模型 License Apache-2.”.
90. GLM-4-9B Split from the former "glm-4" record, which was named for the family while holding one variant's figures. Evidence: "This report primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B." Capabilities evidenced from the primary sources: multilingual: “7k GLM-4 series: Open Multilingual Multimodal Chat LMs | 开源多语言多模态对话模型 License Apache-2.”.
91. Mixtral 8x7B Cited to the lab's own announcement. Not verified because the context window and the parameter count could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.
92. Llama 1 Cited to the lab's own announcement. Not verified because the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from.
93. GLM-130B Cited to the official repository. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Capabilities evidenced from the primary sources: multilingual: “For smaller models, please find monolingual GLMs (English: 10B/2B/515M/410M/335M/110M, Chinese: 10B/335M) and an 1B multilingual GLM (104 languages).”.
94. BLOOM Cited to the official model card. Not verified because the release date could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Capabilities evidenced from the primary sources: multilingual: “0 bloom feature-extraction Model card Files Files and versions Community 23 Train Deploy Use in Transformers Edit model card BLOOM LM BigScience Large Open-science Open-a”.
95. OPT-175B Cited to the technical paper and the official repository. Not verified because the release date and the context window could not be found in any cited primary source — the value may be correct, but this dataset cannot yet show where it came from. Modalities recorded as text-to-text. No image, audio or video capability appears anywhere in the primary sources, which labs advertise prominently when present.

See release analytics →