Viral 'leak' of a giant Anthropic AI model is fake, made of random numbers, users find
A 7.72-terabyte upload on Hugging Face claimed to be a 14-trillion-parameter Anthropic model. Volunteers showed its contents were generated from a simple recipe, not trained.
A huge file posted online and shared on X on Nov. 1 as a leaked Anthropic AI model is a fake made of random numbers, according to Hugging Face users who examined it and a follow-up check by AI company CellCog.
An AI model is mostly a giant list of numbers called weights, which are tuned during training until the model can write or answer questions. The 7.72-terabyte upload, on the model-sharing site Hugging Face under the name burner22/numbat, claimed to be a partly trained version of Anthropic's next top-tier "Mythos" model, with about 14 trillion weights.
Within hours, users found that its numbers could be recreated exactly from each section's name using a standard random-number recipe. One audit matched 43 complete sections, 730,368 values. Trained weights can't be produced that way, so the file was never a trained model.
"I think that confirms it is fake," one user wrote. CellCog said it reran the check and that all nine sections it tested matched.
The repository also asks users to run its own code, which CellCog warned against. Running untrusted code from an anonymous source can put a computer at risk. Anthropic had not commented as of CellCog's check, and no Anthropic statement was found.
Source: Hugging Face discussion #13 on burner22/numbat; discussion #11; discussion #19; CellCog, Nov. 1, 2026