Breaking

Claimed Leak of 14-Trillion-Parameter Anthropic Model on Hugging Face Appears to Contain Randomly Generated Weights

November 1, 2026 at 13:57 EST

Security
Foundation Models

A claimed leak of a 14-trillion-parameter Anthropic model on Hugging Face appears to be a fake: according to a report circulating online, parts of the uploaded weights can be regenerated exactly from nothing but their names, which suggests they were produced by a random number generator rather than by training.

Claimed Anthropic leak · Hugging Face

A claimed 14-trillion-parameter Anthropic model looks like a fake

Parts of the uploaded weights can be regenerated exactly from nothing but their names. That suggests the values came from a random number generator, not from training.

14T

Claimed parameter count of the upload

Tensors reproduced bit for bit from their names

Demonstration of the method, not yet independently verified

Statements from Anthropic in the material reviewed

Tensors reproduced from names alone

Each block is one tensor (43 in total) that matched its regenerated values exactly.

Trained weights would not be expected to match values derived from a seeded generator. A full match is the core of the concern.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.