A Mystery 7.72 TB Numbat Checkpoint Claims to Leak Anthropic's Next AI
A Hugging Face upload called burner22/numbat claims to be a mid-pretraining checkpoint of Anthropic's next Mythos-class model, around 14T parameters.
- Hugging Face repo burner22/numbat claims to be a leaked Anthropic Mythos-class mid-pretraining checkpoint.
- Uploader lists ~14T params, 114 layers, hybrid GDN/MLA, 1,224 MoE experts, NVFP4 quantization.
- Repo ships 772 safetensors shards (~7.72 TB) plus a full reference
modeling_numbat.pyloader. - Hugging Face's automatic counter reads 6.85T parameters, not the claimed 14T.
- Community audits show 43 tensors regenerate bit for bit from name seeded random number generators.
- The burner22 account was created a week before the upload; Anthropic has not commented.
A Hugging Face user named burner22/numbat has published 772 safetensors shards totaling about 7.72 TB. Its model card describes the files as a mid-pretraining checkpoint from Anthropic’s next “Mythos-class” run, with roughly 14 trillion parameters and the internal codename Numbat.
An authentic checkpoint of this scale could expose valuable model weights, architecture choices, and training practices. Anthropic has not publicly commented on the repository, and the available evidence does not establish an Anthropic connection or a data breach. The checkpoint’s provenance remains unverified.
What landed on Hugging Face
| Item | Repository details |
|---|---|
| Files | 772 safetensors shards totaling about 7.72 TB |
| Claimed model | Approximately 14 trillion parameters across 114 layers |
| Automatic count | Hugging Face reports 6.8 trillion parameters. |
| Training point | Step 1,830,000 of an estimated 1.96 million steps |
| Training state | Pretraining checkpoint with no post-training or RLHF |
| Status | Unverified; no public confirmation from Anthropic |
The repository presents itself as a loadable reference implementation. It includes configuration_numbat.py, modeling_numbat.py, a 131,072-token byte-level BPE tokenizer, a shard index, checkpoint metadata, and part of a training log.
The accompanying Python code defines NumbatForCausalLM. The uploader says every tensor matches the module tree, allowing the checkpoint to load with load_state_dict(strict=True) or
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
