7.72TB Hugging Face Repository Emerges; Analysts Suspect Some Weights Are Synthetic

A 7.72TB repository appeared on Hugging Face on October 23, claiming to contain files related to Anthropic's next-generation Mythos-class model, but an analysis published by Reddit user johnxreturn at 1:37 a.m. KST on the 2nd found evidence that some of the weights may be synthetic. The analyst examined portions of the split files, using Qwen2.5-7B as a comparison baseline, and concluded that while the repository's packaging format, tokenizer, and word embeddings showed traces of a real model, the other weights examined did not appear to be genuine training results. The finding matters because a leak of this scale—purporting to involve a frontier lab's unreleased model—would be significant for the AI sector, and questions about its authenticity bear directly on whether any real weights are exposed. Watch for further independent verification of the remaining unexamined shards and any response from Anthropic or the repository owner.