Alleged Mythos Class Anthropic Checkpoint Leak

A Hugging Face repository presents “numbat” as an alleged Anthropic checkpoint from a Mythos-class model’s pretraining run, linking its codename to a phrase said to appear in a March 2026 Claude Code source leak. The uploader describes a hybrid architecture combining Gated DeltaNet and Multi-head Latent Attention, mixture-of-experts layers with 1,224 experts, and multi-token prediction. The model card advertises roughly 14 trillion parameters, while Hugging Face’s metadata reports about 6.8 trillion. Its files are said to span 772 safetensors shards totaling about 7.72 TB, with configuration, tokenizer, tensor index, metadata, training-log excerpts, and a custom reference implementation.

The uploader says the checkpoint is from step 1,830,000 of a planned 1.96 million-step run and has not undergone post-training or RLHF, so its outputs would be raw pretraining completions. The included code is claimed to load the tensors strictly and support inference; the uploader says the checkpoint was quantized to NVFP4 for evaluation, while full-precision training weights and optimizer state stayed within Anthropic’s cluster. To support authenticity, the repository cites consistency checks on tensor formats, quantization values, file sizes, and index entries, along with the alleged codename reference and architectural fit. These are the uploader’s claims; the repository provides no independent confirmation of the checkpoint’s origin.