A Hugging Face repository, identified as 'burner22/numbat', has allegedly posted an internal training checkpoint for Anthropic's Mythos-class model. The purported leak includes details such as 14 trillion parameters, 1,224 Mixture-of-Experts (MoE) experts, and 7.72 TB of weights. The repository also reportedly contains model architecture code, configurations, and training metadata. However, there has been no independent verification of the authenticity of these files, leaving the legitimacy of the leak uncertain. AI
IMPACT Potential leak of a large-scale model could provide insights into advanced AI architectures and training methodologies.
RANK_REASON The cluster discusses an alleged leak of a model's training checkpoint, which falls under research/model release.
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
