A leaked checkpoint for Anthropic's unreleased "Numbat" model, internally codenamed Mythos 5, has surfaced. The checkpoint, totaling approximately 7.72 TB across 772 shards, is a mid-pretraining artifact from a large-scale training run. It includes a complete reference implementation of the model's architecture, which features a hybrid Gated DeltaNet (GDN) and Multi-head Latent Attention (MLA) stack with Mixture-of-Experts layers. AI
IMPACT Potential early access to Anthropic's next-generation model architecture and capabilities.
RANK_REASON Leaked checkpoint of a next-generation model from a frontier lab. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →