What happened
The burner22/numbat card presents an NVFP4 checkpoint across 772 shards (~7.72 TB), hybrid GDN/MLA with 1,224 experts, and says it is Anthropic’s next Mythos-class run under the internal codename numbat. Hugging Face lists the model size as 6.8T params. On November 1, auditor FloxUnder posted discussion #13 with a reproducible recipe: SHA-256 of the tensor name seeds NumPy PCG64; norm weights use exp(0.06 * z); conv kernels use z/6; convert to BF16. The post reports exact matches for 28 normalization tensors (238,848 values) and 15 convolution kernels (491,520 values), then 21 held-out predictions that also matched via HTTP Range reads. The auditor states limits: not every tensor was checked, frozen-parameter training is not logically excluded, and Anthropic origin is not authenticated.
Why it matters
Open-weight discovery still needs the same habit as a leaderboard: name who measured what. A viral “leak” with a matching architecture story can still be synthetic tensors. The useful number here is the audit’s exact match count, not the uploader’s parameter claim.
Hugging Face’s displayed parameter count (~6.8T) disagrees with the card’s ~14T claim; the card attributes the packing to NVFP4.
The auditor published a minimal numpy+requests script that downloads about 90 KB of ranges and asserts exact BF16 matches for two example tensors.
Commenters on the same thread reported some tensors that do not match this recipe, which the thread treats as a possible second recipe or added noise — not as proof of training.