Hugging Face repository claiming to hold a 14-trillion-parameter Anthropic model is a fake built from name-seeded random weights

A repository on Hugging Face that spread on November 1 as a leaked mid-training checkpoint of Anthropic's next Mythos-class model is fake. The 7.72 TB file, posted under the account burner22 as burner22/numbat, contains no trained weights. Its numbers are random, seeded from each tensor's own name, and anyone can regenerate them bit for bit from a few kilobytes of header.

Hugging Face users posted that recipe within two hours of the leak taking off on X. A rerun at 20:18 UTC matched all 9 tensors checked in the first shard - 147,968 values - exactly. A wrong seed matched none. Anthropic had not commented by 20:20 UTC, and its news page lists nothing about the repository.

What the repository claims

The README describes a model of "~14T param, 114 layers," a mixture of experts with 1,224 experts (48 active per token plus one shared), stored in NVFP4, a 4-bit format. It calls numbat Anthropic's internal codename and points to a comment it says sat in the March 2026 Claude Code source leak: "Remove this section when we launch numbat." That comment has not been verified, and a codename, even a real one, says nothing about whose weights these are.

The repository was created on October 23, 2026 at 01:37 UTC by burner22, and its files last changed on October 24. The README ends with "Mirrors welcome. Do not DM me asking for the source." By 20:03 UTC on November 1, Hugging Face's counter showed 5,846 downloads and 162 likes.

The surface details hold up better than the weights. The 772 shards and 7.72 TB are real, but the content is generated. The NVFP4 quantization is internally consistent, which is beside the point: generated numbers quantize cleanly too. A commenter found the separate output head equals 0.891 times the embedding table. Another listed bugs in the reference code: no KV cache, a no-op weight init, and NaN on padded rows.

How the fake was found

On November 1 at 14:53 UTC, @OxShamss posted the repository as one of the craziest AI leaks ever, while noting that nobody had verified the weights. The post passed 231,000 views by 20:03 UTC. A second account posted "ALL 7.72TB SECURED" a little after 17:00 UTC. On Hugging Face, the checking had already started.

At 16:03 UTC a user posted a code review of the model files and called the release a well made fake. At 16:29 UTC a user named FloxUnder posted an audit: 43 complete tensors, 730,368 values, could be regenerated exactly from their names with a short script that downloads about 90 KB. Within an hour others had reproduced it, one writing "I think that confirms it is fake." At 19:43 UTC a user named cekal extended the recipe to the large 4-bit matrices, which make up most of the file, and reported "162,246,496 out of 162,246,496 matches" on signs across 5,503 tensors from 770 of the 772 shards.

We reran the check

The recipe is short. Read the first 8 bytes of shard 1 to get the header length, then the header itself: 16,344 bytes of JSON listing every tensor's name, shape and byte range. For each tensor, hash its name with SHA-256, take the first 8 bytes as a number, and use it to seed NumPy's PCG64 random generator. Draw normal random values in the tensor's shape, scale them, round to bfloat16, and compare with the bytes on the server.

Six normalization tensors from layers 0 and 1 matched exactly, along with three convolution kernels from layer 0. The control did not: the same tensor with seed 12345 failed to match. Only the header and those tensors were downloaded, using HTTP range requests, and none of the repository's code was run. A trained model cannot come out of a random generator whose only input is the tensor's name.

How to check a leaked checkpoint before you download it

• Read the header with a range request first. Never pull terabytes to answer a yes-or-no question.
• Test whether small tensors regenerate from a simple seed such as the name or the index. Trained norms and kernels never do.
• Do not load custom code from an anonymous repository. This one ships its own modeling file and asks for remote code to be trusted.
• Run the economics. One commenter wrote, "This model is impossible to be real due to serving economics by any standard."
• Look for any first-party trace: a vendor page, a model card, a statement. Here there was none.

These habits apply well beyond one hoax. Teams evaluating open-weight models are increasingly asked to make calls on partial information, and the same verification steps - header first, small tensors second, custom code last - apply to any download. For anyone building on models rather than auditing them, courses on generative AI and LLMs cover how weights, formats and checkpoints actually work.

Why someone would build it

Nobody knows. Commenters guessed a prank and, in one case, an attempt to move prices ahead of Anthropic's IPO; there is no evidence for either. What is clear is the care: matching configuration files, a README section arguing why readers should believe it, and 4-bit blocks that round-trip cleanly. One commenter noticed that "they even added noise to the initialized weights that could make a naive test assume they were actually trained."

The repository also says nothing about Mythos 5.5. It predates the real signal of the day: Anthropic's own Mythos page showed a hidden Mythos 5.5 label in an October 31 archive copy, then read 5.1 again on November 1.

Why this matters for creatives, developers and comms teams

A 7.72 TB file with a polished README can fool a lot of people in a few hours. If your job involves citing model releases, writing about them, or building on them, the header check is a two-minute task that settles the question. Ask for a first-party source before you repeat a leak, and treat download counts and likes as popularity, not proof. The next fake will likely be better made than this one.

Sources: Hugging Face repository burner22/numbat, README, config and model API, revision 6b7f8fc5, read November 1, 2026, 20:03 UTC; Hugging Face discussions #11, #13, #14 and #19, read November 1, 2026, 20:17 UTC; X post by @OxShamss, November 1; NumPy PCG64 generator; rerun of the discussion #13 recipe on shard 1, November 1, 2026, 20:18 UTC.