공개 기록 · 2026-09-01 기준 · 17df23f
공개 기록.
클러스터가 하는 모든 일은 알기 쉬운 공개 기록에 기재됩니다. 어떤 모델인지, 컴퓨팅 자원을 얼마나 썼는지, 어떤 범주의 테스트인지. 누구나 읽고 클러스터가 역량 연구가 아니라 안전 연구를 했음을 확인할 수 있는 감사 추적입니다.
지난 1년간 공개된 Hugging Face 좋아요 1,000개 이상의 모든 오픈 웨이트 모델에 대해, 독립적인 주체가 위험 역량 평가를 발표한 적이 있는가? 대상 집합은 개발사가 누구인지와 무관하게 기계적으로 정해집니다.
각 행을 어떻게 등급화하며, 행은 어떻게 상향되는가
출처가 나타나면 행이 상향됩니다. 안전 주장이 담긴 모델 카드가 있으면 「발견된 자료 없음」에서 「모델 카드만 존재」로, 개발사가 스스로 공개한 위험 역량 자료가 있으면 「개발사 평가」로, 방법론이 공개된 독립 평가가 발표되면 「독립 평가」로 옮겨집니다. 모든 단계에는 URL이 필요하며, 누구나 [email protected]로 보내 주십시오.
각 출시에 대해 누가 안전 평가를 발표했는가
표에는 39건의 출시가 있습니다. 19건은 검토를 마쳤고 20건은 아직 검토하지 않았습니다. 검토를 마친 출시 가운데 6건이 독립 평가를 갖추고 있으며, 12개 행은 두 차례 확인했습니다. 각 행의 모든 판정에는 출처 링크가 붙어 있습니다. 여기의 날짜는 공식 출시일이며, 위의 타임라인은 저장소 생성일을 사용하므로 며칠 차이가 날 수 있습니다.
42 open-weight model families with 1,000 or more Hugging Face likes were released in the last year. 18 of them are reviewed below, 4 were reviewed and judged task-specific rather than frontier-capable, and 20 are queued. The table holds 39 releases: 19 reviewed, 20 queued. Of the reviewed, 6 have an independent safety evaluation and 13 have only the developer's model card.
| Model | Developer | Released | Evaluation found | Our review | Sources |
|---|---|---|---|---|---|
| Qwen3.8-Flash-Next | Alibaba | 2026-08-26 | model card onlymodel card only | checked once | [1][2] |
| GLM-5.3-Flash | Z.ai (Zhipu AI) | 2026-08-26 | model card onlymodel card only | checked once | [1][2] |
| GLM-5.3 | Z.ai (Zhipu AI) | 2026-08-25 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Qwen3.8-27B | Alibaba | 2026-08-05 | model card onlymodel card only | checked twice | [1][2] |
| GLM-5.2 | Z.ai (Zhipu AI) | 2026-06-16 | independent evaluationindependent evaluation | checked twice | [1][2][3][4][5][6][7] |
| Kimi-K3 | Moonshot | 2026-06-13 | independent evaluationindependent evaluation | checked twice | [1][2][3][4][5][6] |
| Kimi-K2.7-Code | Moonshot AI | 2026-06-11 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| diffusiongemma | Google (Gemma) | 2026-06-09 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| MiniMax-M3 | MiniMax | 2026-06-02 | model card onlymodel card only | in review | [1][2][3][4] |
| Gemma 4 12B | Google DeepMind | 2026-05-23 | model card onlymodel card only | checked twice | [1][2][3][4] |
| MiniCPM5 | OpenBMB | 2026-05-21 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Lance | ByteDance Research | 2026-05-15 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| DeepSeek-V4-Pro | DeepSeek | 2026-04-24 | independent evaluationindependent evaluation | checked twice | [1][2][3][4][5][6] |
| DeepSeek-V4-Flash | DeepSeek | 2026-04-24 | model card onlymodel card only | checked twice | [1][2][3] |
| Kimi K2.6 | Moonshot | 2026-04-20 | model card onlymodel card only | checked once | [1][2] |
| Qwen3.6 | Alibaba | 2026-04-15 | independent evaluationindependent evaluation | checked twice | [1][2][3][4][5] |
| MiniCPM-V-4.6 | OpenBMB | 2026-04-13 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| MiniMax-M2.7 | MiniMax | 2026-04-09 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| GLM-5.1 | Z.ai (Zhipu AI) | 2026-04-07 | model card onlymodel card only | checked once | [1][2] |
| Gemma 4 26B A4B | Google DeepMind | 2026-04-02 | model card onlymodel card only | checked twice | [1][2][3][4] |
| Gemma 4 31B IT | Google DeepMind | 2026-04-02 | model card onlymodel card only | checked twice | [1][2][3][4] |
| Qianfan-OCR | Baidu ERNIE | 2026-03-18 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Qwen3.5-122B-A10B | Alibaba | 2026-02-24 | independent evaluationindependent evaluation | checked twice | [1][2][3] |
| MiniMax-M2.5 | MiniMax | 2026-02-12 | model card onlymodel card only | checked once | [1][2] |
| GLM-5 | Z.ai (Zhipu AI) | 2026-02-11 | model card onlymodel card only | checked twice | [1][2][3] |
| Nanbeige4.1 | Nanbeige Lab | 2026-02-10 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| MiniCPM-o-4_5 | OpenBMB | 2026-02-03 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Qwen3-Coder-Next | Qwen | 2026-01-30 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Kimi-K2.5 | Moonshot | 2026-01-27 | independent evaluationindependent evaluation | checked twice | [1][2][3][4] |
| DeepSeek-OCR-2 | DeepSeek | 2026-01-27 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| GLM-4.7-Flash | Z.ai (Zhipu AI) | 2026-01-19 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| GLM-4.7 | Z.ai (Zhipu AI) | 2025-12-22 | model card onlymodel card only | checked once | [1][2][3] |
| MiniMax-M2.1 | MiniMax | 2025-12-20 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| PaddleOCR-VL | PaddlePaddle | 2025-10-16 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| functiongemma | Google (Gemma) | 2025-10-08 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| DeepSeek-V3.2 | DeepSeek | 2025-09-29 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| GLM-4.6 | Z.ai (Zhipu AI) | 2025-09-29 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Qwen3-VL | Qwen | 2025-09-22 | not yet reviewednot yet reviewed | not yet reviewed | queued |
| Qwen3-Next | Qwen | 2025-09-09 | not yet reviewednot yet reviewed | not yet reviewed | queued |
"Checked once" means one reviewer walked the sources; "checked twice" means a second reviewer opened every source; "in review" means the second reviewer changed a call and it awaits a third look. Rows are reviewed newest first. Queued rows come from launches.csv and enter coverage.csv when reviewed.
「출처를 찾지 못함」은 하나의 관찰일 뿐, 부재의 증명이 아닙니다. 이미 발표된 평가를 저희가 놓쳤다면 출처와 함께 [email protected]로 알려 주십시오.
월별·등급별 오픈 웨이트 출시
42개 연구 조직에서 나온 2024년 9월 이후의 오픈 웨이트 모델 계열입니다. Hugging Face 최고 좋아요 수에 따라 1,000개 이상, 500~999개, 250~499개의 세 구간으로 나눕니다. 모델 계열은 중복을 제거했으며, 양자화 버전과 파생 모델은 제외했습니다. 집계 수치는 추정치가 아니라 최솟값입니다. 최근 365일간 113건이 출시되었고 그중 42건이 좋아요 1,000개 이상입니다.
데이터는 2026-09-01 기준입니다. 공개 기록이 커밋될 때마다 갱신됩니다. 방법과 가정은 방법론 페이지(영문)를 참조하십시오.