“Numbat” Is Not a Model

Last updated: November 7, 2026

A 7.7 TB Hugging Face upload claims to be a leaked Anthropic training checkpoint. We sampled its weights. They are noise in the right file format.

The Hugging Face repo burner22/numbat is not a trained language model. Its card presents it as a mid-pretraining checkpoint of an unreleased Anthropic model: about 14 trillion parameters, 772 files, 7.7 TB. We read the weights directly, without downloading the full repo or running its code. The two ends of the model, where words go in and come out, look like a real model. Every matrix we sampled between them has the statistics of random noise. Don’t spend the bandwidth, and don’t load it with trust_remote_code.

What the repo claims

  • The claim. The card says the files are an NVFP4 copy of Anthropic’s “next Mythos-class training run”, codename numbat, taken at step 1,830,000 before any post-training.
  • The account. It was created a week before the upload and has published nothing else. The repo has drawn over 9,000 downloads and 370 likes in two weeks.
  • The evidence offered. The card’s case is that every file is internally consistent: the 4-bit blocks decode correctly and the file sizes agree with the index. That is true, and it shows nothing. Any program that converts random numbers to NVFP4 properly passes the same checks.
  • Why few people can check. Running it would take about 7.7 TB of accelerator memory, so almost no one can test the claim by using the model. The weights themselves can still be tested.

What the weights show

We read individual rows from seven of the 772 files, spread from the first to the last, and measured the shape of each weight matrix: how heavy its tails are, and whether a few directions dominate it. We ran the same measurements on two real trained models, Qwen3-30B-A3B and Qwen3-14B, and on random numbers, to check the test tells them apart.

  • Trained weights are uneven. In the real models, tail heaviness varies more than thirtyfold from one matrix to the next. The router, which picks which experts handle each token, is dominated by a couple of directions.
  • Numbat’s weights are all the same. Every interior matrix we sampled, from layer 0 to the last, has the same profile: dense layers, attention, the router and the experts. It is the profile of random noise drawn from a Laplace distribution, and a noise sample with light scaling matches it closely.
  • This is a sample, not a full audit. We tested 12 interior matrices out of more than 1.5 million tensors. But they span the whole depth of the model and every kind of layer, and not one looked trained.
  • A noise router can’t work. Numbat’s router has no dominant directions at all, so it would spread tokens across experts essentially at random. Even setting provenance aside, these files could not function as a mixture-of-experts language model.

The two ends look real

The input embeddings, which turn tokens into vectors, and the output layer, which turns vectors back into tokens, both contain meaning: “cat” sits near “dog”, “king” near “queen”, “Paris” near “London”, at almost the same strength in both. These are the first places a skeptic would look, so a quick spot check there passes. They are also only a small part of the weights. Everything that would have to do the thinking in between is noise.

The rest of the story doesn’t hold up either

  • The design is assembled from open models. It borrows DeepSeek-V3’s attention dimensions and expert routing, and the attention mix of Qwen3-Next and Kimi-Linear. None of that is evidence of an unreleased lab design.
  • The training log implies too much data. Its throughput and step count imply roughly 160 trillion training tokens, several times the largest pretraining run any lab has disclosed.
  • The format story is odd. The card says the evaluation copy was stored in NVIDIA’s NVFP4 format to run on Amazon’s Trainium2 chips, which as far as we know don’t support it.

What we changed

Our free scanner checks a model’s configuration against the ModelAtlas reference, which is fast but can’t see weights. On this repo it had matched only a vocabulary size and rated it clean. We have added this weight-level finding to it: scanning burner22/numbat on modeldna.ai now returns a high-risk warning.

The general point: a model card’s internal consistency is not provenance. File formats are easy to satisfy. What a model learned is in the statistics of its weights, and those can be checked without trusting anything the uploader wrote.

Sources: the repo’s card, configuration, checkpoint metadata and training-log tail (as of November 7, 2026); weight rows read by HTTP range from shards 1, 23, 150, 400, 650, 771 and 772. Controls: Qwen3-30B-A3B and Qwen3-14B, read the same way, and synthetic Laplace noise.