burner22/numbat: Leaked pretraining checkpoint for Anthropic's numbat model
An unreleased mid-pretraining checkpoint for Anthropic's next Mythos-class model, codenamed numbat, has been leaked on Hugging Face. The ~14T parameter model features a hybrid Gated DeltaNet and Multi-head Latent Attention architecture with 1,224 Mixture-of-Experts layers and is quantized in NVFP4. The repository includes a complete reference implementation and 772 safetensors shards totaling approximately 7.72 terabytes of tensor data. Open article →
Notable comments
The leak is widely considered fake, with commenters pointing to “name-seeded RNG” artifacts, partial tensor matches to NumPy outputs, and implausibility around how/why such a large checkpoint would remain hosted for a week without takedown.
HN · kingstnap Claims the weights are random: “matrixes are literally just random data and the seed is the tensor name,” implying fabricated checkpoint contents.
HN · free_bip References specific technical evidence—“many tensors are a bit-perfect match to NumPy RNG output”—and notes the practical cost of verifying a 14T run.
HN · jackfischer Raises a credibility/takedown red-flag: asks whether it’s realistic the repo could stay on Hugging Face for a week without being removed for copyright.
- ‼️ Someone "leaked" a 14-trillion-parameter Anthropic model on Hugging Face, but it contains weig...
- The alleged Anthropic Numbat leak is likely fake. Parts of the model match randomly generated num...
- Alleged Mythos Class Anthropic Checkpoint Leak
- The "Mythos leak" on Hugging Face is not real. I went through the repo myself. It was uploaded b...
- > The checkpoint's architecture (hybrid GDN/MLA, 1,224 experts, NVFP4) matches what would be e...
- Anthropic's Mythos-class model allegedly leaked on Hugging Face