AI BriefingAudit reveals alleged 14T Anthropic 'numbat' checkpoint is seeded random noise, not a trained model
·2026.11.03 16:30
Key point
A viral 7.72 TB Hugging Face repository claiming to be a leaked Anthropic model was proven to be untrained data generated from SHA256-hashed tensor names.
Details
- The 'numbat' repository, circulating as a leaked 14T parameter Anthropic checkpoint, contains weights that are bit-for-bit identical to random noise generated by seeding NumPy PCG64 with SHA256 hashes of tensor names.
- Audits confirmed 162,246,496 of 162,246,496 sign bits matched across 770 of 772 shards, and a control test with a wrong seed matched only about 0.6% of values.
- Statistical analysis shows norm layer kurtosis is approximately 3 at every depth, whereas trained models like Qwen3 and DeepSeek V3 exhibit kurtosis between 20 and 150.
- The repository's README instructs users to load the model with trust_remote_code=True, a security risk given the file is not a valid trained model.
- The 'numbat' codename originates from a March Claude Code leak, which likely contributed to the initial spread of the fake checkpoint.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Anthropic amended S-1 reveals $17 billion Q3 revenue and $895.4 billion in compute commitments
Training 1,000 Weights in Gemma 3 270M Unlocks Hidden Sorting and JSON SkillsAnthropic restricts Claude from reverse-engineering video games