Leak of Anthropic's Secret Mythos Model
Files attributed to Anthropic's secret Mythos model have appeared on the Hugging Face platform. The total size of the uploaded data is 7.72 terabytes.
The claimed neural network architecture includes 14 trillion parameters and 114 layers. It uses a MoE mechanism involving 1,224 experts, allowing only a subset of the system to be activated for each query.
The model is still in the training phase and has not undergone final tuning or Reinforcement Learning from Human Feedback (RLHF). Researchers have already identified potential errors in the model's weights.
Automatic parameter detection shows a value of 6.8 trillion, although the author claims 14 trillion. Earlier in March, the source code for Claude Code was leaked.
An unknown user posted the files on Hugging Face without explanation. According to available data, more than five thousand people have already downloaded this data.










