Repository Claims Anthropic Checkpoint Leak

An anonymous Hugging Face repository called “numbat” claims to contain a checkpoint from an Anthropic Mythos-class training run, pulled during pretraining. The page lists about 14T parameters, 114 layers, a hybrid GDN/MLA architecture, 1,224 experts with top-48 routing plus 1 shared expert, and NVFP4 storage. The repository’s YAML metadata is empty or missing.

The repository lists 772 safetensors shards totaling 7.72 TB, with each file under 10.6 GB. It includes configuration and model code, a tokenizer with 131,072 tokens, an index covering all shards, checkpoint metadata, and the last ~400 training steps. The author says the included reference implementation has a matching module for every tensor and can load the checkpoint with strict=True or through AutoModelForCausalLM using trust_remote_code=True. The model’s forward method runs the trunk, while its draft method runs the Multi-Token Prediction layer.

The uploader argues the files are credible because “numbat” appeared as the codename of an unreleased Anthropic model in a Claude Code source leak from March 2026, and because the architecture and tensor data are internally consistent. The page says NVFP4 scale blocks decode within RTNE bounds, scales do not underflow, and each NVFP4 quad round-trips exactly; shard sizes, tensor counts, and index entries are also described as self-consistent. These are claims on the repository page, not independent confirmation of Anthropic’s involvement.

The uploader says the checkpoint is at step 1,830,000 of an approximately 1.96M-step training run, with no post-training or RLHF, and describes its output as raw pretraining completion. They say evaluation copies are quantized to NVFP4 to run on about 85 Trainium2 chips; the BF16 training weights and optimizer state remain inside the cluster. The page reports 5,846 downloads last month, lists a model size of 6.8T parameters, and says no inference provider hosts it.

An anonymous Hugging Face repository called “numbat” claims to contain a checkpoint from an Anthropic Mythos-class training run, pulled during pretraining. The page lists about 14T parameters, 114 layers, a hybrid GDN/MLA architecture, 1,224 experts with top-48 routing plus 1 shared expert, and NVFP4 storage. The repository’s YAML metadata is empty or missing.

The repository lists 772 safetensors shards totaling 7.72 TB, with each file under 10.6 GB. It includes configuration and model code, a tokenizer with 131,072 tokens, an index covering all shards, checkpoint metadata, and the last ~400 training steps. The author says the included reference implementation has a matching module for every tensor and can load the checkpoint with strict=True or through AutoModelForCausalLM using trust_remote_code=True. The model’s forward method runs the trunk, while its draft method runs the Multi-Token Prediction layer.

The uploader argues the files are credible because “numbat” appeared as the codename of an unreleased Anthropic model in a Claude Code source leak from March 2026, and because the architecture and tensor data are internally consistent. The page says NVFP4 scale blocks decode within RTNE bounds, scales do not underflow, and each NVFP4 quad round-trips exactly; shard sizes, tensor counts, and index entries are also described as self-consistent. These are claims on the repository page, not independent confirmation of Anthropic’s involvement.

The uploader says the checkpoint is at step 1,830,000 of an approximately 1.96M-step training run, with no post-training or RLHF, and describes its output as raw pretraining completion. They say evaluation copies are quantized to NVFP4 to run on about 85 Trainium2 chips; the BF16 training weights and optimizer state remain inside the cluster. The page reports 5,846 downloads last month, lists a model size of 6.8T parameters, and says no inference provider hosts it.