Anthropic / Wikimedia Commons (Public domain)
A 7.72 TB upload billed as a Mythos-class checkpoint appears to contain tensors generated from their own names
A Hugging Face repository called “numbat” claims to hold a mid-pretraining checkpoint from Anthropic’s upcoming Mythos-class training run. If real, it would be one of the most significant AI leaks on record.
Community auditors say it probably isn’t real. Their analysis suggests at least part of the upload was manufactured from scratch rather than lifted from an Anthropic data center.
What the numbat repo claims to be #
The repository sits under the username burner22, an account created on October 16, 2026. The upload followed a week later, on October 23, 2026.
The model described in the repo is enormous. It reportedly packs approximately 14 trillion parameters spread across 114 layers.
The claimed architecture also mixes a hybrid GDN/MLA design with 1,224 Mixture of Experts (MoE) components. MoE is a setup where the model is carved into many specialist sub-networks, and only a handful switch on for any given input.
The files themselves total around 7.72 TB across 772 safetensors shards. Safetensors is a common file format for storing model weights, chosen partly because it is safer to load than older formats.
The upload also includes configuration files and a byte-level BPE tokenizer with a vocabulary of 131,072 tokens.
Why auditors think the weights are fake #
The most damaging finding came from a community analysis dated November 1, 2026. Auditors found 43 tensors, containing 730,368 BF16 values, that regenerate exactly from their own tensor names.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
The recipe works like this: take a tensor’s name, run it through a SHA-256 hash, and use the result to seed a random number generator. The output matches the stored values precisely.
Reruns of the test reportedly confirmed the matches were exact, not approximate. That precision is what moved the conversation from “suspicious” to “likely synthetic.”
The parameter count adds another wrinkle. The repo advertises roughly 14 trillion parameters, while Hugging Face’s own metadata lists approximately 6.8 trillion.
Analysts also flagged the overall weight distributions. According to community analyses, they look more like generated noise than the patterns expected from trained models such as Qwen2.5.
Why the ‘numbat’ name carries baggage #
The codename itself is not random. “Numbat” was previously linked to a March 2026 leak of Claude Code source, which exposed a set of internal names.
It also cuts the other way. Anyone could have picked the name up from coverage of the earlier leak, so its presence proves familiarity with public reporting rather than access to Anthropic’s servers.
Interest in the repo has been substantial anyway. By November 1, it had logged about 5,846 downloads and more than 162 likes.
Anthropic has not issued a public statement on the repository or the claims attached to it.
What this means for AI labs, platforms and developers #
For Hugging Face, the case raises an awkward point about scale. An anonymous week-old account was able to host 7.72 TB of disputed files and attract thousands of downloads before the strongest counter-evidence surfaced. What to watch next is whether further audits extend the synthetic-tensor finding beyond the first 43 tensors. A match across more of the 772 shards would turn “evidence suggests” into something closer to a closed case.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.