@kqyno7: Anthropic's Mythos-class model may have just been uploaded straight to Hugging Face? A repo called burner22/numbat claims to have obtained an Anthropic internal training checkpoint, with a mind-blowing scale: • 14T (
Summary
A Hugging Face repo called burner22/numbat claims to contain an Anthropic internal training checkpoint, reportedly 14T parameters, 114 layers, and 1,224 MoE experts — suspected to be one of the largest leaks of closed-source model weights ever.
View Cached Full Text
Cached at: 11/01/26, 05:56 PM
💀 Anthropic’s Mythos-class model, allegedly uploaded straight to Hugging Face?
A repo named burner22/numbat claims to have obtained an internal Anthropic training checkpoint, and the scale is pretty absurd:
- 14T (14 trillion) parameters, 114 layers
- 1,224 MoE experts, with 48 active per token
- 7.72 TB of weights, split across 772 Safetensors files
- A GDN + MLA hybrid architecture with MTP (multi-token prediction) support
- NVFP4 quantization, reportedly from mid-pretraining
Taking a look at the repo, it does indeed list the model config, architecture implementation, training metadata, and a large number of weight shards — not just an empty README.
The author also claims that “numbat” is an internal codename that previously appeared in the Claude Code source code leak.
If this is real, it could be one of the most stunning closed-source model weight leaks in history.
Real leak, or a deliberately designed fake model? 👀
#Anthropic #Claude #Mythos #HuggingFace #LLM #AI
Similar Articles
burner22/numbat
A Hugging Face repo publishes a claimed leaked mid-pretraining checkpoint of Anthropic's unreleased ~14T-parameter Mythos-class model codenamed "numbat" (referenced in a Claude Code source leak), including 772 safetensors shards (~7.72 TB), a full reference implementation, and training logs at step 1.83M of ~1.96M.
@AYi_AInotes: Mind blown—22-year-old founder open-sources what Anthropic kept locked away: the Claude Mythos black box
A 22-year-old founder claims to have cracked open the architecture black box of Anthropic’s Claude Mythos model via an open-source project, speculating it uses a recurrent-depth Transformer design instead of simply scaling parameters.
Alleged Mythos Class "leak" is fake. Forensic Investigation inside.
A forensic analysis concludes that the alleged Anthropic 'numbat' 14T-parameter checkpoint on Hugging Face is fake: packaging, tokenizer, and embedding matrix are genuine, but every weight tensor is synthetic, proven by an lm_head/embedding linear relationship and inconsistent NVFP4 input_scale calibration values.
The alleged 14T Anthropic "numbat" checkpoint on Hugging Face doesn't look trained: its weights regenerate exactly from name seeded random noise
An alleged 14T-parameter Anthropic "numbat" checkpoint leaked on Hugging Face turns out to be seeded random noise: its weights regenerate exactly from SHA256-hashed tensor names, with 162M+ sign bits matching across 770 of 772 shards, and README asks users to load it with trust_remote_code=True.
AI News: The Model That Has Everyone Freaked Out!
Anthropic’s unreleased Claude Mythos model demonstrates elite-level hacking and vulnerability discovery, prompting a private preview with major tech firms to patch software before public release.