A community audit reproduced 43 tensors bit-for-bit, and an independent check reproduced nine samples, supporting the conclusion that at least the checked weights are synthetic noise.
Community audit finds the claimed Anthropic Numbat weights are seeded noise
2026-11-02Frontier watchIGNORE
A Hugging Face repository claimed to contain a multi-terabyte checkpoint from an unreleased Anthropic model. A reproducible community audit regenerated 43 complete tensors exactly from a pseudorandom generator seeded by tensor names, and an independent check reproduced nine sampled tensors. Anthropic has not authenticated the repository, so the upload is not evidence of a model leak or release.
What this means for you
For teams monitoring open weights, popularity and internally consistent file layouts are not provenance. This item should be ignored as a model signal and treated as a reminder to verify origin and sample weights before mirroring or running custom code.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Uploader Claim | Unreleased Anthropic Mythos-class checkpoint; not authenticated by Anthropic |
|---|---|
| Repository Size | Uploader claims 772 shards totaling about 7.72 TB |
| Reproducible Check | 43 complete tensors totaling 730,368 BF16 values matched name-seeded NumPy RNG output exactly |
| Independent Check | 9 of 9 sampled tensors, 147,968 values, reproduced exactly; most large NVFP4 tensors were outside that independent sample |
| Safety Boundary | Repository includes custom modeling code; no full download or code execution was needed for the range-read audit |
| Time Basis | Audit discussion opened 2026-11-01T16:29:50Z inside the overlap; stored as a late discovery |
Technical impact 1
Engineering value 2
Adoption signal 3
Maturity 1
Verification cost 2
Risk 4