A Hugging Face upload presented as a leaked Anthropic checkpoint contains 43 tensors that can be rebuilt exactly from their own names with a seeded random number generator, according to an audit posted on the repository’s community page on 1 November. The upload, burner22/numbat, claims about 14T parameters; Hugging Face’s automatic count shows 6.85T.
The audit’s author, a Hugging Face user named FloxUnder, wrote that “43 complete tensors — 730,368 BF16 values — can be regenerated bit for bit from their tensor names alone”. By our arithmetic, those values are about one in 9.4 million of the 6.85T Hugging Face counts.
Business Pill · Least Privilege
A short explainer on the least-privilege principle. It teaches the general idea only and says nothing about this upload.
The key insight: As we read it, the community audit is a test that needs no trust in the uploader: it checks small byte ranges against a published recipe, and its author still limits what the match proves.
FloxUnder’s post, discussion #13 on the repository, says the 43 tensors are 28 normalization weights and 15 convolution kernels, and that every one of their 730,368 values matched numbers generated from a seed derived from each tensor’s name.
The post says predictions were recorded for 21 tensors before they were read, and that “Every predicted value matched the subsequent HTTP Range reads exactly.” It adds: “Only small byte ranges were downloaded, and no Python code from the model repository was executed.”
The post itself sets limits: “These matches alone cannot logically exclude a training setup that deliberately kept all these parameters frozen.” It adds: “They also do not authenticate the claimed Anthropic origin.”
As we read it, the finding shows these 43 tensors match a random number generator exactly; the audit covers 43 tensors, not the whole upload.
Hugging Face model API and model card for burner22/numbat (read 1 November 2026): the uploader claims about 14T parameters; Hugging Face counts 6.85T, about 6.83T of it as 8-bit unsigned integers. Two 4-bit values per 8-bit slot would give about 13.67T (our arithmetic, given the card’s NVFP4 claim).
What the Uploader Claims
The model card describes a “~14T param” model with 114 layers and 1,224 experts, and calls it “Anthropic’s next Mythos-class training run”, a checkpoint “pulled mid-pretraining”.
It ties the name to Anthropic’s March Claude Code source leak, citing a comment in prompts.ts: “Remove this section when we launch numbat.” The card also says: “Mirrors welcome. Do not DM me asking for the source.”
Hugging Face’s user record shows the burner22 account was created on 16 October and the repository on 23 October. By our arithmetic, that is seven days. The API listed 5,846 downloads and 137 likes when we read it.
The Parameter Count Gap
Another user, Metris, flagged the count gap in discussion #6: “Hugging Face automatically reports 6.8T parameters, while the uploader claims approximately 14T.”
Hugging Face’s count lists about 6.83T of the 6.85T as 8-bit unsigned integers. The card says the weights are stored in NVFP4, a 4-bit format. By our arithmetic, two 4-bit values per 8-bit slot would give about 13.67T, close to the claim, so the count gap alone does not settle the question.
A third user, shadowlilac, argued in discussion #14 that “This model is impossible to be real due to serving economics by any standard.”
Why It Matters
As we read it, the upload shows how a codename the card ties to an earlier leak can lend credibility to a multi-terabyte file that few people can test. The card says the 772 shards total about 7.72 TB and that loading uses the repository’s own model code.
The audit took the other route: it read small byte ranges and ran no code from the repository. As we read it, that difference matters most for anyone who would load the files with the repository’s own code.
The Structural Read
As we read it, the upload leans on a codename that its own model card ties to Anthropic’s March Claude Code source leak; the sources here do not establish any link to Anthropic.
By our arithmetic, the 730,368 regenerated values are about one in 9.4 million of the 6.85T parameters Hugging Face counts, so the audit is narrow but exact.
As we read the repository’s discussions, users raised three separate doubts: the parameter count, serving economics and the random-number tensors.
FloxUnder, Hugging Face discussion #13, 1 November 2026
“can be regenerated bit for bit from their tensor names alone”
Three Implications
FOR ANYONE MIRRORING LARGE UPLOADS FloxUnder says the audit downloaded only small byte ranges and executed no Python code from the repository, while the model card says loading uses the repository’s own model code.
FOR MODEL HUBS Hugging Face’s public records show the account was created on 16 October and the repository on 23 October, and the API listed 5,846 downloads when we read it.
FOR AI LABS The model card presents the upload as an Anthropic checkpoint; FloxUnder writes that the matches do not authenticate the claimed Anthropic origin.
The framework says: “Every technology stack has layers. Some are scarce, some are abundant.”
As we read it, trained frontier weights sit in the scarce layer, which is why a claimed leak draws attention. The framework predates this upload and does not discuss it.
The sources here are the repository, its model card and its community posts. They do not establish who uploaded it, whether any part of it came from Anthropic, or what the rest of the weights contain. The audit tested 43 tensors; the conclusions about the remainder are not in it.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
A user audit found that 43 tensors in the “14T” numbat upload match a seeded random number generator exactly, while its author notes the result does not by itself rule out deliberately frozen parameters. As we read it, the parameter-count gap has a possible 4-bit explanation, by our arithmetic; the audit’s finding is the stronger signal.
A note on sourcing. Figures come from Hugging Face’s public model and user records, the uploader’s model card and the repository’s community discussions, read on 1 November. We did not download the weights or run any code from the repository, and we did not contact the uploader, the auditor, Anthropic or Hugging Face. Comparisons labelled as our arithmetic are ours.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
