Know what a number means.

01

Release tracking

The snapshot queries 42 Hugging Face organisation IDs and reads 6,977 repositories. It retains text-generation, image-text-to-text and any-to-any models with at least 250 likes, excluding named quantizations and derivative formats.

Families merge parameter-size, precision, checkpoint and specified release suffixes. Product tiers such as Flash, Pro, Coder and Vision stay separate. The earliest repository creation date is used as a release-date proxy, and the highest repository like count sets the popularity band.

These are nested thresholds, not three additive counts. Likes measure popularity, not capability. New releases have had less time to accumulate likes; ModelScope-only releases are not scanned; family matching is heuristic. The monthly chart excludes the incomplete as-of month.

02

Evaluation coverage

The review follows model cards, technical reports, release posts and independent evaluation reports. Each reviewed entry receives its highest supported public-evidence tier. The highest tier does not describe every evaluation the model has had.

T3Independent evaluation

An independent dangerous-capability evaluation with a public methodology. This does not imply comprehensive testing or a safe model.

T2Developer evaluation

A developer-run dangerous-capability evaluation. The developer is not an independent evaluator.

T1Developer documentation

A model card or ordinary safety disclosure, without a qualifying dangerous-capability evaluation found in the source review.

T0No evaluation found

A completed source review found no qualifying evaluation or safety disclosure. Not proof of absence.

—Not yet reviewed

The review has not been completed. Never counted as T0 or as a missing evaluation.

Why 42 families and 40 entries?

Of 42 high-popularity families, 17 are represented, 4 are below the frontier-scope rule, and 21 await review. The 19 coded records include 18 release-specific entries and one Qwen3.6 family entry; Gemma 4 has several checkpoint records. MiniMax-M2 is now correctly queued: it must not inherit MiniMax-M2.5’s review.

“Independent evaluation” describes the model evaluator: outside the developer, with a published method and a relevant result. “Internal second pass” describes our record review: 12 of 19 coded entries received a second AI-assisted source check on 1 September. Neither label means a third party audited this dataset.

V2 rechecked the tier-deciding reports for all six T3 records and checked all 57 cited URLs in the coded records. It corrected a family-match error, a date and a redirected source. This remains an internal review. An external auditor has not signed off; incomplete negative findings and checkpoint ambiguities remain visible.

03

Cost estimates

The current evaluation price is $200k all-in, including delivery staff, compute, storage and transfer, investigation and scoped retesting, and allocated operations and contingency. It was set on September 7, 2026. It is a programme price, not a provider quote, measured average or externally audited cost model.

04

Key terms

Open weights

Model parameters available to download and run. Access and licence terms vary; “open weights” is not the same as unrestricted open source.

Dangerous-capability evaluation

Testing whether a model can materially enable serious harm—for example in biological, cyber or autonomous tasks. A test covers a defined scope, not every possible use.

Black-box / white-box

Black-box testing uses inputs and outputs, usually through an endpoint. White-box work also accesses internal model information, such as weights or activations.

Campaign

A scoped programme of evaluation, replication and follow-up around a model. Not the same unit as a launch family, a ledger entry or a GPU job.

05

Download & cite

The downloads contain public fields and source links, not internal coding notes. Keep the snapshot date and methodology with any reused figure.

• Codebook & field definitions ↓
• Evidence records · CSV ↓
• Source-check receipt · JSON ↓
• Campaign assumptions · JSON ↓

To propose a correction, send the model, the claim to change and a primary source to [email protected].