This repo already has ~6k downloads, so surely those who really care (OpenAI, DeepSeek, etc) have a copy without doing the leg work themselves (assuming it is real). Although, I thought at this point, the pre-training between the companies would be mostly identical, so all of the real work comes afterwards.
https://huggingface.co/burner22/numbat/discussions/13
This comment even has the receipts.
Idk how his AI figured this out, but the matrixes are literally just random data and the seed is the tensor name.
My own agent reproduced the findings.
[0]: https://www.reddit.com/r/LocalLLaMA/comments/1xl7xdk/is_that...
explosion-s•1h ago
Not legit. The repo is real and the shards are real files, but they are not an Anthropic checkpoint. The only evidence on the model card is the name numbat. That is a real internal codename from the public Claude Code source leak ("remove this section when we launch numbatiại). That comment does not describe this architecture, and anyone can read it. The card contradicts itself. It says the weights were quantized to NVFP4 so evals could run on Trainium2. Trainium2 does not support NVFP4. AWS documents its types as FP32, TF32, BF16, FP16, and FP8. NVFP4 is an NVIDIA format. The architecture is also just public papers stapled together: Gated DeltaNet, DeepSeek-style latent attention, 1,224 experts, multi-token prediction. Nothing ties that stack to Anthropic. A few smaller tells: the account was created on Oct 16 and is named burner, the readme says ~141 params while the file index counts 6.85T, and the training log only ever advances in steps of 6, 7, or 8 seconds, including straight through a line that claims a 178-second checkpoint. It has been public since Oct 23. Anthropic DMCA'd the Claude Code mirrors within days. This is still up. Don't run the custom Python in that repo.