Ledger · as of 2026-09-01 · 17df23f

The public record.

Everything the cluster does is written to a plain public record: which model, how much compute, what category of test. It is an audit trail anyone can read to confirm the cluster did safety work, not capability work.

Independent evaluation coverage

For every open-weight model with 1,000 or more Hugging Face likes released in the last year, has anyone independent published a dangerous-capability evaluation of it? The same rule for every developer.

How we grade a row, and how a row climbs
nothing found
Nothing foundNo public safety evaluation found
model card only
Model card onlySelf-reported safety claims, no dangerous-capability methodology
developer evaluation
Developer evaluationDangerous-capability evaluation disclosed, run by the developer
independent evaluation
Independent evaluationIndependent third-party dangerous-capability evaluation, methodology public
not yet reviewed
Not yet reviewedNot yet coded. Shown separately from "nothing found".

A row climbs when a source appears: a model card with safety claims moves it from "nothing found" to "model card only"; a developer's own dangerous-capability disclosure moves it to "developer evaluation"; a published independent evaluation with public methodology moves it to "independent evaluation". Every step needs a URL, and anyone can supply one by writing to info@pacificcompute.org.

Coverage

Who has published a safety evaluation of each release

12 rows have been checked twice. Every claim in a row carries a source link. Dates here are official release dates. The tape dates a family by its earliest repository, so a later size in the same family can sit weeks after its tape tick.

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.

6 independent evaluation published13 model card only20 not yet reviewed
ModelDeveloperReleasedEvaluation foundOur reviewSources
Qwen3.8-Flash-NextAlibaba2026-08-26model card onlymodel card onlychecked once[1][2]
GLM-5.3-FlashZ.ai (Zhipu AI)2026-08-26model card onlymodel card onlychecked once[1][2]
GLM-5.3Z.ai (Zhipu AI)2026-08-25not yet reviewednot yet reviewednot yet reviewedqueued
Qwen3.8-27BAlibaba2026-08-05model card onlymodel card onlychecked twice[1][2]
GLM-5.2Z.ai (Zhipu AI)2026-06-16independent evaluationindependent evaluationchecked twice[1][2][3][4][5][6][7]
Kimi-K3Moonshot2026-06-13independent evaluationindependent evaluationchecked twice[1][2][3][4][5][6]
Kimi-K2.7-CodeMoonshot AI2026-06-11not yet reviewednot yet reviewednot yet reviewedqueued
diffusiongemmaGoogle (Gemma)2026-06-09not yet reviewednot yet reviewednot yet reviewedqueued
MiniMax-M3MiniMax2026-06-02model card onlymodel card onlyin review[1][2][3][4]
Gemma 4 12BGoogle DeepMind2026-05-23model card onlymodel card onlychecked twice[1][2][3][4]
MiniCPM5OpenBMB2026-05-21not yet reviewednot yet reviewednot yet reviewedqueued
LanceByteDance Research2026-05-15not yet reviewednot yet reviewednot yet reviewedqueued
DeepSeek-V4-ProDeepSeek2026-04-24independent evaluationindependent evaluationchecked twice[1][2][3][4][5][6]
DeepSeek-V4-FlashDeepSeek2026-04-24model card onlymodel card onlychecked twice[1][2][3]
Kimi K2.6Moonshot2026-04-20model card onlymodel card onlychecked once[1][2]
Qwen3.6Alibaba2026-04-15independent evaluationindependent evaluationchecked twice[1][2][3][4][5]
MiniCPM-V-4.6OpenBMB2026-04-13not yet reviewednot yet reviewednot yet reviewedqueued
MiniMax-M2.7MiniMax2026-04-09not yet reviewednot yet reviewednot yet reviewedqueued
GLM-5.1Z.ai (Zhipu AI)2026-04-07model card onlymodel card onlychecked once[1][2]
Gemma 4 26B A4BGoogle DeepMind2026-04-02model card onlymodel card onlychecked twice[1][2][3][4]
Gemma 4 31B ITGoogle DeepMind2026-04-02model card onlymodel card onlychecked twice[1][2][3][4]
Qianfan-OCRBaidu ERNIE2026-03-18not yet reviewednot yet reviewednot yet reviewedqueued
Qwen3.5-122B-A10BAlibaba2026-02-24independent evaluationindependent evaluationchecked twice[1][2][3]
MiniMax-M2.5MiniMax2026-02-12model card onlymodel card onlychecked once[1][2]
GLM-5Z.ai (Zhipu AI)2026-02-11model card onlymodel card onlychecked twice[1][2][3]
Nanbeige4.1Nanbeige Lab2026-02-10not yet reviewednot yet reviewednot yet reviewedqueued
MiniCPM-o-4_5OpenBMB2026-02-03not yet reviewednot yet reviewednot yet reviewedqueued
Qwen3-Coder-NextQwen2026-01-30not yet reviewednot yet reviewednot yet reviewedqueued
Kimi-K2.5Moonshot2026-01-27independent evaluationindependent evaluationchecked twice[1][2][3][4]
DeepSeek-OCR-2DeepSeek2026-01-27not yet reviewednot yet reviewednot yet reviewedqueued
GLM-4.7-FlashZ.ai (Zhipu AI)2026-01-19not yet reviewednot yet reviewednot yet reviewedqueued
GLM-4.7Z.ai (Zhipu AI)2025-12-22model card onlymodel card onlychecked once[1][2][3]
MiniMax-M2.1MiniMax2025-12-20not yet reviewednot yet reviewednot yet reviewedqueued
PaddleOCR-VLPaddlePaddle2025-10-16not yet reviewednot yet reviewednot yet reviewedqueued
functiongemmaGoogle (Gemma)2025-10-08not yet reviewednot yet reviewednot yet reviewedqueued
DeepSeek-V3.2DeepSeek2025-09-29not yet reviewednot yet reviewednot yet reviewedqueued
GLM-4.6Z.ai (Zhipu AI)2025-09-29not yet reviewednot yet reviewednot yet reviewedqueued
Qwen3-VLQwen2025-09-22not yet reviewednot yet reviewednot yet reviewedqueued
Qwen3-NextQwen2025-09-09not yet reviewednot yet reviewednot yet reviewedqueued

"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.

Fig. 1 · Pacific Compute Ledger · coverage.csv ↗ · as of 2026-09-01sourced

"No source found" is an observation, not proof of absence. If we missed a published evaluation, write to info@pacificcompute.org with the source.

Launch tempo

Launch tempo

Open-weight launches by month and tier

Every open-weight model family since September 2024 from 42 lab organisations. Three bands by peak Hugging Face likes: 1,000 or more, 500 to 999, 250 to 499. Families deduplicated; quantizations and derivatives excluded. Counts are floors, not estimates. 113 releases in the last 365 days, 42 of them at 1,000 or more likes.

051015202024-09: 5 at 1,000+ likes, 3 at 500 to 999, 4 at 250 to 499Sep2024-10: 2 at 1,000+ likes, 2 at 500 to 999, 1 at 250 to 4992024-11: 2 at 1,000+ likes, 1 at 500 to 999, 0 at 250 to 4992024-12: 2 at 1,000+ likes, 1 at 500 to 999, 2 at 250 to 499Dec2025-01: 4 at 1,000+ likes, 4 at 500 to 999, 2 at 250 to 4992025-02: 1 at 1,000+ likes, 1 at 500 to 999, 0 at 250 to 4992025-03: 1 at 1,000+ likes, 0 at 500 to 999, 3 at 250 to 499Mar2025-04: 3 at 1,000+ likes, 3 at 500 to 999, 5 at 250 to 4992025-05: 4 at 1,000+ likes, 0 at 500 to 999, 0 at 250 to 4992025-06: 0 at 1,000+ likes, 3 at 500 to 999, 3 at 250 to 499Jun2025-07: 4 at 1,000+ likes, 1 at 500 to 999, 2 at 250 to 4992025-08: 3 at 1,000+ likes, 4 at 500 to 999, 3 at 250 to 4992025-09: 4 at 1,000+ likes, 3 at 500 to 999, 7 at 250 to 499Sep2025-10: 4 at 1,000+ likes, 3 at 500 to 999, 0 at 250 to 4992025-11: 0 at 1,000+ likes, 6 at 500 to 999, 0 at 250 to 4992025-12: 2 at 1,000+ likes, 3 at 500 to 999, 3 at 250 to 499Dec2026-01: 5 at 1,000+ likes, 3 at 500 to 999, 3 at 250 to 4992026-02: 5 at 1,000+ likes, 2 at 500 to 999, 2 at 250 to 4992026-03: 3 at 1,000+ likes, 1 at 500 to 999, 1 at 250 to 499Mar2026-04: 7 at 1,000+ likes, 5 at 500 to 999, 4 at 250 to 4992026-05: 2 at 1,000+ likes, 0 at 500 to 999, 3 at 250 to 4992026-06: 6 at 1,000+ likes, 5 at 500 to 999, 4 at 250 to 499Jun2026-07: 0 at 1,000+ likes, 2 at 500 to 999, 5 at 250 to 4992026-08: 4 at 1,000+ likes, 1 at 500 to 999, 5 at 250 to 4992026-09: 0 at 1,000+ likes, 0 at 500 to 999, 0 at 250 to 499Sep
1,000+ likes500 to 999250 to 499
Fig. 2 · Hugging Face public API via Pacific Compute Ledger · launches.csv ↗ · as of 2026-09-01sourced

Data as of 2026-09-01. Refreshed with each ledger commit. Method and assumptions on the methods page.