Unreleased Anthropic Model Checkpoint Leaked Online
A machine learning repository has surfaced containing a massive, unreleased model checkpoint codenamed “numbat,” which is reportedly linked to Anthropic’s next generation of Mythos-class technology. The files appeared on the Hugging Face platform, providing a substantial look into the architecture of a training run that was previously referenced in earlier code leaks.
The leaked data consists of 772 shards totaling approximately 7.72 terabytes of tensor information. According to the provided documentation, the model is a hybrid structure utilizing Gated DeltaNet (GDN) and Multi-head Latent Attention (MLA), incorporating a Mixture-of-Experts (MoE) configuration with 1,224 experts. The repository includes the necessary Python files, specifically a reference implementation that allows users to load the checkpoint using standard machine learning tools.
The codename “numbat” previously appeared in a March 2026 leak of the Claude Code source, where it was mentioned in internal comments regarding future product launches. Technical analysis of the leaked files shows that the checkpoint represents a model mid-pretraining, specifically at step 1,830,000 of a planned 1.96 million-step process. It remains a raw pretraining output, lacking post-training refinement or Reinforcement Learning from Human Feedback (RLHF).
The files are presented in an NVFP4 quantized format. The repository owner explains that the evaluation team typically quantizes checkpoints to this format to facilitate testing on specific computing clusters, such as Trainium2 hardware, rather than requiring the massive infrastructure used during the initial training phase. The repository includes strict consistency across all shards, confirming the legitimacy of the tensor data architecture and internal decoding scales.
While the uploader has declined to provide details regarding the origin of the data, they have indicated that the shards are currently being uploaded for public access. The inclusion of modeling_numbat.py enables users to perform inference using the model, despite the incomplete nature of the pretraining run. As of late October 2026, the repository has attracted significant interest, with thousands of downloads and interactions on the platform. The leak appears to be a notable event in the landscape of AI development, offering an unprecedented look at the technical architecture of high-end, unreleased training runs.
Sources: Hacker News
Aggregated summary. Original: Hacker News