Artificial Intelligence1 min read

Alleged Mythos Class Anthropic Checkpoint Leak

By · Published by Everything Blog

In short

The publicly shared repository contains an evaluation copy captured at step 1,830,000 of an estimated 1.96-million-step pretraining run under the internal codename "numbat." To facilitate testing outside full training clusters, the massive model was quantized using the NVFP4 format across 772 safetensors files. The drop includes complete reference architecture code, tokenizer assets, and training logs to support direct execution. It incorporates a trunk forward pass alongside a dedicated Multi-Token Prediction layer.

Key points

  • The model lacks post-training or reinforcement learning from human feedback because it wa…: The model lacks post-training or reinforcement learning from human feedback because it was captured directly from pretraining.
  • The "numbat" codename previously surfaced in a March 2026 Claude Code source leak within…: The "numbat" codename previously surfaced in a March 2026 Claude Code source leak within a prompts file comment.
  • Specialized hardware slices can load and execute the model using the provided reference c…: Specialized hardware slices can load and execute the model using the provided reference code and quantized tensor weights.

Sources checked by our analysis pipeline: 1

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

numbat

~14T param, 114 layers, hybrid GDN/MLA, MoE (1,224 experts, top-48 + 1

shared), NVFP4. Internal codename numbat

, Anthropic's next Mythos-class

training run. numbat

is the codename referenced in the Claude Code

source leak (Mar 2026). This is a checkpoint from that training run,

pulled mid-pretraining.

Files:

File | Contents |

---|---|

config.json |

model configuration |

configuration_numbat.py |

NumbatConfig |

modeling_numbat.py |

NumbatForCausalLM reference implementation |

tokenizer.json |

byte-level BPE, 131,072 tokens |

model.safetensors.index.json |

tensor index across all 772 shards |

model-00001-of-00772.safetensors |

shard 1 of 772 |

ckpt_metadata.json |

checkpoint metadata (ckpt path, eval harness output) |

training_log_tail.txt |

last ~400 steps of training log |

The 772 shards total ~7.72 TB of tensor data. Each shard is safetensors format, under 10.6 GB per file.

configuration_numbat.py

and modeling_numbat.py

are included: a

complete reference implementation of the checkpoint. Its module tree

has an exact slot for every tensor in every shard, so the full

checkpoint loads with load_state_dict(strict=True)

(and, since…

Artificial Intelligence

More in Artificial Intelligence