Alleged Anthropic 'Numbat' checkpoint on Hugging Face draws authenticity challenges

A Hugging Face repository called numbat presents itself as a checkpoint from what its owner calls Anthropic's next “Mythos-class” training run, taken partway through pretraining. That account comes from the repository owner, not from Anthropic.

The repository card claims roughly 14 trillion parameters, 114 layers and 772 weight shards totaling about 7.72 TB. But the same page's Hugging Face interface displays a model size of 6.8 trillion parameters. Asked when more shards would be uploaded, the card answers, “They're all uploading. Be patient.” Its description does not establish how many shards were available.

On Nov. 1, 2026, X user challenged the card's assertion that its described architecture matches what one would expect of a Mythos-class run: “no, it's not what would be expected.” In a separate , kimmonismus called the alleged leak “likely fake.” The user asserted that parts of the model match randomly generated numbers exactly, suggesting they were never trained. The post provides no measurements or method for that assertion. The repository's provenance claim and the two users' challenges leave the checkpoint's authenticity unresolved.