Alleged Anthropic 'Numbat' model checkpoint posted online in claimed leak

A repository posted online claims to contain a mid-pretraining checkpoint of Anthropic's next large-scale model, internally codenamed Numbat — described as the company's next Mythos-class training run.

The posting claims the model has approximately 14 trillion parameters across 114 layers, using a hybrid Gated DeltaNet and Multi-head Latent Attention architecture with a Mixture-of-Experts feed-forward layer containing 1,224 experts. The checkpoint data totals roughly 7.72 terabytes across 772 shards in safetensors format.

The poster states the checkpoint was pulled mid-pretraining at step 1,830,000 of an approximately 1.96 million-step run, with no post-training or reinforcement learning from human feedback applied. The weights are stored in NVFP4 format, which the poster says is the format used by Anthropic's evaluation team to run assessments on a slice of Trainium2 hardware.

The posting includes configuration and modelling files described as a complete reference implementation. The codename Numbat is cited as matching a reference in a Claude Code source code leak reported in March 2024.

Anthropic has not publicly confirmed or commented on the posting. The poster declined to disclose the source of the material.