A Hugging Face post claims to contain an Anthropic Mythos-class checkpoint

A Hugging Face Blog post alleges that a large, mid-training checkpoint is from an Anthropic model codenamed numbat. The post describes its architecture and files, but the supplied evidence contains no independent confirmation of the claim.

• The Hugging Face post describes the alleged checkpoint as a roughly 14-trillion-parameter model with 114 layers and a hybrid Gated DeltaNet and Multi-head Latent Attention architecture.· 1 source
• The post says the checkpoint is at training step 1,830,000 of an approximately 1.96-million-step run and has not undergone post-training or RLHF.· 1 source
• The post claims the 772 shards contain about 7.72 TB of tensor data; the Hacker News evidence only provides a title and says the item contains comments.· 2 sources