The Pulse
Hugging Face Upload Billed as a 14T Anthropic Checkpoint Was Seeded Noise
A Hugging Face repository advertised as a 14-trillion-parameter Anthropic checkpoint drew scrutiny after users showed its tensors could be regenerated from their names using seeded random-number recipes. Follow-up checks reproduced values i
AI.info Team · Written with AI · fact-checked ·
A 7.72-terabyte upload, and a fast-moving claim
A Hugging Face repository presented as a roughly 14-trillion-parameter Anthropic checkpoint drew attention on November 1, then faced a technical challenge: auditors found that its tensors could be recreated from their names using a seeded random-number generator. CellCog, which published its own check that evening, reported that the tested values matched exactly. The evidence points to synthetic data rather than trained model weights; it does not establish who uploaded the files or why.
The repository, named burner22/numbat, described itself as a mid-training checkpoint for Anthropic’s next Mythos-class model. Its card claimed about 14 trillion parameters, 114 layers and a mixture-of-experts design, and listed 772 file shards totaling about 7.72 terabytes. The uploader also called “numbat” an Anthropic codename, but CellCog said it had not verified the cited reference to an earlier Claude Code source leak.
Tensor names reproduced the values
The first detailed audit, posted in the repository’s Hugging Face discussion on November 1, reported that 43 complete normalization and convolution tensors—730,368 values—could be regenerated bit for bit. The recipe hashes each tensor’s name with SHA-256, uses the first eight bytes to seed NumPy’s PCG64 generator, then applies specified transformations and converts the results to the format stored in the files. The auditor said the script fetched roughly 90 kilobytes and did not run code supplied by the repository.
In the discussion, FloxUnder, the Hugging Face user who posted the audit, wrote:
“I checked this repository through an AI-assisted, reproducible audit and found an exact deterministic generation fingerprint in the published weights.”
— FloxUnder, Hugging Face user and audit author
CellCog reported a separate check of the first shard: all nine tensors tested matched, covering 147,968 values, while a control using a different seed did not.
Follow-up checks reached the large matrices
The initial audit had a limit: it tested a subset of the files, not every tensor, and its author noted that those matches alone could not rule out a hypothetical training setup that deliberately froze the tested parameters. A later Hugging Face discussion extended the analysis to the large 4-bit matrices. Its author reported matching the signs of 162,246,496 values across 5,503 tensors in 770 of the 772 shards.
That follow-up also rebuilt ten matrices from layers 2 through 108 and reported that 465,566,255 of 465,567,744 four-bit codes matched. Those checks give the case more reach than the initial sample, although they are not a complete byte-for-byte reconstruction of every tensor in the repository. CellCog’s conclusion that the upload is not a trained checkpoint rests on these reproducible patterns, rather than on the repository’s size, configuration or unusual implementation details.
The upload does not authenticate an Anthropic model
The repository’s listed files included custom modeling code, and CellCog advised readers not to run code from an anonymous upload. The presence of plausible architecture files and internally consistent quantized data cannot establish that Anthropic created or trained the weights; the tensor checks instead provide evidence that much of the content was generated by a deterministic recipe.
CellCog reported that Anthropic had not commented on the repository by 20:20 UTC on November 1. It also said there was no first-party vendor page, model card or statement corroborating the leak. The repository therefore offered no verified link to Anthropic, while the public tensor audits gave independent readers a compact way to test what its files contained.