A pretraining checkpoint from Anthropic's next-generation Mythos-class model, "numbat", has been leaked, containing a hybrid GDN/MLA architecture with 1,224 experts and NVFP4. The model's architecture and internal consistency match expectations for a Mythos-class model, and a reference implementation is included in the repository. The checkpoint is from a mid-pretraining stage, with 772 shards and a total size of approximately 7.72 TB. AI summary
Firehose
Filtered to Hacker News, tagged “mythos” · clear filters
Browse: People · Companies · Papers · Podcasts · Hacker News · Deep dives