Alleged Anthropic Mythos model leak on Hugging Face is likely fake, Reddit users say
A Hugging Face repository claiming to contain training files for an Anthropic "Mythos-class" model spread across Reddit on Nov. 1, drawing roughly 2,670 upvotes on the main post. After examining the files, users said the supposed 7.72 TB checkpoint does not appear to be a real Anthropic model. Their view was that, aside from the tokenizer, packaging, and word embedding matrix, most of the weights were synthetically generated from random noise. The most detailed analysis came from Reddit user johnxreturn, who estimated there was only a 0.5% chance the repository contained a genuine Anthropic checkpoint. He said the file structure looked convincing and that the claimed parameter count broadly checked out after accounting for a format that Hugging Face may not have counted. But he argued the actual weight content failed multiple consistency tests, including abnormal similarity between output and embedding layers, mismatched quantization calibration values, and weight distributions that did not resemble those of a trained model. Other users also flagged contradictions in the repository documentation and warned against running bundled custom Python code. Anthropic had not publicly responded at the time of publication.
