Leaked Anthropic "numbat" checkpoint surfaces on Hugging Face
Alleged Mythos Class Anthropic Checkpoint Leak
A Hugging Face repo posted by user burner22 hosts what it claims is a mid-pretraining checkpoint from Anthropic's next Mythos-class run, codenamed numbat. The listing describes a ~14T-parameter model with 114 layers, hybrid GDN/MLA attention, 1,224 MoE experts, and NVFP4 quantization, spread across 772 safetensors shards totaling about 7.72 TB, with a full reference implementation that loads via strict state dict. The uploader ties the codename to a comment in the March 2026 Claude Code source leak and offers no further explanation of provenance.
This is a pretraining checkpoint at step 1,830,000 of a ~1.96M step run. No post-training, no RLHF. Raw pretraining-completion output. That's expected.