Alleged 14T Anthropic "numbat" leak debunked: weights regenerate from name-seeded random noise
djcroman · reddit · 2026-11-03
The burner22/numbat repo circulating on X and HN as a leaked 14T-param Anthropic mid-training checkpoint (7.72 TB, 772 shards) falls apart under a simple test: hashing tensor names with SHA256 to seed NumPy's PCG64 and drawing normals reproduces the stored weights bit-for-bit. Audits matched 43 full tensors plus 162,246,496/162,246,496 sign bits across the 4-bit expert matrices; a control with a wrong seed matched only 0.6%. Norm-layer kurtosis (3) looks nothing like trained models (Qwen3, DeepSeek V3: 20–150). The "numbat" codename is real from the March Claude Code leak, but Anthropic hasn't commented. Note: the README asks for trustremotecode=True, a potential risk, and the author asks whether tools exist to auto-flag seeded-noise checkpoints.
Related event: Alleged Anthropic 14T-Parameter Checkpoint Leak Proven Fake(3 posts)→
More from Models
- Unverified leak claims Fable 5.5 generates full animation in one shot, beating Opus 5.5 — FinanceYF5 · 2026-11-03
- User tests Grok's new order tracking: it reports coffee bean delivery status — jonathan_wilke · 2026-11-03
- Is Sonnet really worse than Opus at task execution? Reddit debates model tiers — Top-Lab9125 · 2026-11-03
- Alignment debate: treat AI as allies—slow down and stop deceiving them — davidmanheim · 2026-11-03
- Altman: Next GPT Solved Navier-Stokes With 10,000 Agents Working 4 Days Straight — DeryaTR_ · 2026-11-03
- DeepSeek V5 reportedly planned to be pre-trained on Huawei Ascend chips — hedgehog0 · 2026-11-03