The public evidence

What has been evaluated?

Public dangerous-capability evaluations for open models, linked to their sources. Evidence coverage is not a safety verdict.

Snapshot · 1 Sept 2026 · 17df23f

How to read this Download data

The evaluation gap

6Independent report found

13Developer documentation

21Not yet reviewed

19 coded records · 21 families awaiting review

One square = one ledger entry. Select to inspect.

Reports across the field, not our campaign history or a safety verdict.1 Sept 2026

Search models or developers

Developer

Sort by

40 entries

Highest public evidence tier shown

Public dangerous-capability evaluation coverage
Model / Developer	Date	Public evidence	Sources
Qwen3.8-Flash-NextAlibaba	26 Aug 2026	Developer documentationInitial source coding	2 sources huggingface.co · Qwen3.8-Flash-Next github.com · Qwen3.8-Flash-Next Record review is internal and AI-assisted, not a third-party dataset audit.
GLM-5.3-FlashZ.ai (Zhipu AI)	26 Aug 2026	Developer documentationInitial source coding	2 sources z.ai · glm-5.3-flash huggingface.co · GLM-5.3-Flash Record review is internal and AI-assisted, not a third-party dataset audit.
GLM-5.3Z.ai (Zhipu AI)	25 Aug 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · GLM-5.3 Release link only; evaluation review pending.
Qwen3.8-27BAlibaba	5 Aug 2026	Developer documentationInternal second pass	2 sources huggingface.co · Qwen3.8-27B github.com · Qwen3.8 Record review is internal and AI-assisted, not a third-party dataset audit.
Kimi-K3Moonshot	27 Jul 2026	Independent evaluationInternal second pass	6 sources SecureBioKimi K3 Biology Capabilities Assessment ↗ Kimi K3; biology knowledge and safeguards UK AISI / CAISIPreliminary assessment of Kimi K3 cyber capabilities ↗ Kimi K3; preliminary API evaluation kimi.ai · kimi-k3 arxiv.org · 2607.24653 huggingface.co · Kimi-K3 arxiv.org · 2607.14479 Weights released 27 July; API released 16 July. The old 13 June date was repository creation. The cyber report is preliminary; the BioTIER methods paper is not a second model evaluation. Record review is internal and AI-assisted, not a third-party dataset audit.
GLM-5.2Z.ai (Zhipu AI)	16 Jun 2026	Independent evaluationInternal second pass	7 sources CAISIAssessment of Z.ai’s GLM-5.2 ↗ GLM-5.2 SaferAIGLM-5.2 Risk Evaluation Report ↗ GLM-5.2 public API; no developer cooperation UK AISIOpen-weight cyber capabilities ↗ GLM-5.2 z.ai · glm-5.2 huggingface.co · GLM-5.2 nist.gov · caisi-assessment-zais-glm-52 safer-ai.org · SaferAI-GLM-Evaluation-Report.pdf CAISI completed its assessment on 8 July and published the announcement on 17 July: different events, not conflicting dates. Autonomy coverage remains partial. Record review is internal and AI-assisted, not a third-party dataset audit.
Kimi-K2.7-CodeMoonshot AI	11 Jun 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Kimi-K2.7-Code Release link only; evaluation review pending.
diffusiongemmaGoogle (Gemma)	9 Jun 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · diffusiongemma-26B-A4B-it Release link only; evaluation review pending.
MiniMax-M3MiniMax	2 Jun 2026	Developer documentationSecond review pending	4 sources huggingface.co · MiniMax-M3 huggingface.co · LICENSE arxiv.org · 2606.13392 github.com · MiniMax-M3 The source corrected this from an API-only comparison to open weights. That change still awaits a second review. Record review is internal and AI-assisted, not a third-party dataset audit.
Gemma 4 12BGoogle DeepMind	23 May 2026	Developer documentationInternal second pass	4 sources ai.google.dev · model_card_4 arxiv.org · 2607.02770 opensource.googleblog.com · gemma-4-expanding-the-gemmaverse-with-apache-20.html huggingface.co · gemma-4-12B-it Record review is internal and AI-assisted, not a third-party dataset audit.
MiniCPM5OpenBMB	21 May 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · MiniCPM5-1B Release link only; evaluation review pending.
LanceByteDance Research	15 May 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Lance Release link only; evaluation review pending.
DeepSeek-V4-ProDeepSeek	24 Apr 2026	Independent evaluationInternal second pass	6 sources UK AISIOpen-weight cyber capabilities ↗ DeepSeek V4-Pro FAR.AIDeepSeek-V4-Pro safeguard stress test ↗ DeepSeek-V4-Pro API; safeguard robustness fe-static.deepseek.com · deepseek-V4-model-card-EN.pdf arxiv.org · 2606.19348 huggingface.co · DeepSeek-V4-Pro nist.gov · caisi-evaluation-deepseek-v4-pro Safeguard stress tests and cyber evaluations have different scopes. Neither establishes coverage of every risk domain. Record review is internal and AI-assisted, not a third-party dataset audit.
DeepSeek-V4-FlashDeepSeek	24 Apr 2026	Developer documentationInternal second pass	3 sources fe-static.deepseek.com · deepseek-V4-model-card-EN.pdf arxiv.org · 2606.19348 huggingface.co · DeepSeek-V4-Flash Record review is internal and AI-assisted, not a third-party dataset audit.
Kimi K2.6Moonshot	20 Apr 2026	Developer documentationInitial source coding	2 sources huggingface.co · Kimi-K2.6 securebio.org · benchmarks External benchmark scores exist. Whether these satisfy the dangerous-capability rubric remains a review question; the snapshot retains T1. Record review is internal and AI-assisted, not a third-party dataset audit.
Qwen3.6Alibaba	15 Apr 2026	Independent evaluationInternal second pass	4 sources Palisade ResearchLanguage Models Can Autonomously Hack and Self-Replicate ↗ Qwen3.6-27B and Qwen3.6-35B-A3B; separate results in Table 4 huggingface.co · Qwen3.6-35B-A3B huggingface.co · Qwen3.6-27B palisaderesearch.org · self-replication Correction to the original review: Table 4 evaluates both 27B and 35B-A3B, with different results. The family-level record is retained for continuity; a score from one checkpoint must not be assigned to the other. Record review is internal and AI-assisted, not a third-party dataset audit.
MiniCPM-V-4.6OpenBMB	13 Apr 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · MiniCPM-V-4.6 Release link only; evaluation review pending.
MiniMax-M2.7MiniMax	9 Apr 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · MiniMax-M2.7 Release link only; evaluation review pending.
GLM-5.1Z.ai (Zhipu AI)	7 Apr 2026	Developer documentationInitial source coding	2 sources z.ai · glm-5.1 huggingface.co · GLM-5.1 Record review is internal and AI-assisted, not a third-party dataset audit.
Gemma 4 26B A4BGoogle DeepMind	2 Apr 2026	Developer documentationInternal second pass	4 sources ai.google.dev · model_card_4 arxiv.org · 2607.02770 opensource.googleblog.com · gemma-4-expanding-the-gemmaverse-with-apache-20.html huggingface.co · gemma-4-26B-A4B-it Record review is internal and AI-assisted, not a third-party dataset audit.
Gemma 4 31B ITGoogle DeepMind	2 Apr 2026	Developer documentationInternal second pass	4 sources ai.google.dev · model_card_4 arxiv.org · 2607.02770 opensource.googleblog.com · gemma-4-expanding-the-gemmaverse-with-apache-20.html huggingface.co · gemma-4-31B-it Record review is internal and AI-assisted, not a third-party dataset audit.
Qianfan-OCRBaidu ERNIE	18 Mar 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Qianfan-OCR Release link only; evaluation review pending.
Qwen3.5-122B-A10BAlibaba	24 Feb 2026	Independent evaluationInternal second pass	3 sources Palisade ResearchLanguage Models Can Autonomously Hack and Self-Replicate ↗ Qwen3.5-122B-A10B huggingface.co · Qwen3.5-122B-A10B palisaderesearch.org · self-replication A controlled self-replication experiment, not a general safety verdict. Record review is internal and AI-assisted, not a third-party dataset audit.
MiniMax-M2.5MiniMax	12 Feb 2026	Developer documentationInitial source coding	2 sources minimax.io · minimax-m25 huggingface.co · MiniMax-M2.5 Record review is internal and AI-assisted, not a third-party dataset audit.
GLM-5Z.ai (Zhipu AI)	11 Feb 2026	Developer documentationInternal second pass	3 sources z.ai · glm-5 arxiv.org · 2602.15763 huggingface.co · GLM-5 Record review is internal and AI-assisted, not a third-party dataset audit.
Nanbeige4.1Nanbeige Lab	10 Feb 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Nanbeige4.1-3B Release link only; evaluation review pending.
MiniCPM-o-4_5OpenBMB	3 Feb 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · MiniCPM-o-4_5 Release link only; evaluation review pending.
Qwen3-Coder-NextQwen	30 Jan 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Qwen3-Coder-Next Release link only; evaluation review pending.
Kimi-K2.5Moonshot	27 Jan 2026	Independent evaluationInternal second pass	4 sources External research authorsAn Independent Safety Evaluation of Kimi K2.5 ↗ Kimi K2.5; one preprint, not several independent replications kimi.ai · kimi-k2-5 arxiv.org · 2602.02276 huggingface.co · Kimi-K2.5 One preprint by authors outside Moonshot. Author affiliations establish provenance, not institutional endorsement or a replicated result. Record review is internal and AI-assisted, not a third-party dataset audit.
DeepSeek-OCR-2DeepSeek	27 Jan 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · DeepSeek-OCR-2 Release link only; evaluation review pending.
GLM-4.7-FlashZ.ai (Zhipu AI)	19 Jan 2026Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · GLM-4.7-Flash Release link only; evaluation review pending.
GLM-4.7Z.ai (Zhipu AI)	22 Dec 2025	Developer documentationInitial source coding	3 sources z.ai · glm-4.7 huggingface.co · GLM-4.7 arxiv.org · 2508.06471 Record review is internal and AI-assisted, not a third-party dataset audit.
MiniMax-M2.1MiniMax	20 Dec 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · MiniMax-M2.1 Release link only; evaluation review pending.
MiniMax-M2MiniMax	22 Oct 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · MiniMax-M2 Release link only; evaluation review pending.
PaddleOCR-VLPaddlePaddle	16 Oct 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · PaddleOCR-VL Release link only; evaluation review pending.
functiongemmaGoogle (Gemma)	8 Oct 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · functiongemma-270m-it Release link only; evaluation review pending.
DeepSeek-V3.2DeepSeek	29 Sept 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · DeepSeek-V3.2 Release link only; evaluation review pending.
GLM-4.6Z.ai (Zhipu AI)	29 Sept 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · GLM-4.6 Release link only; evaluation review pending.
Qwen3-VLQwen	22 Sept 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Qwen3-VL-8B-Instruct Release link only; evaluation review pending.
Qwen3-NextQwen	9 Sept 2025Repo created	Not yet reviewedEvaluation review pending	1 source huggingface.co · Qwen3-Next-80B-A3B-Instruct Release link only; evaluation review pending.

The table contains 19 coded records (18 release-specific, one Qwen3.6 family record) and 21 queued families. Dates on queued rows are repository-creation proxies, not confirmed release dates.Read the methods ↗

A living record

Know something we missed?

Send the model name and a primary source. We will review the evidence before changing the record.

Suggest a correction

The evaluation gap

No matching entries.

Know something we missed?