Hugging Face verified a 16.5-trillion-parameter model of zeros

Vacuum-16t declares 16,501,264,351,232 parameters across 3,841 tensors on Hugging Face, per its model card.

Every one of those parameters is a zero, and the listing still passes Hugging Face's verification and shows 906 downloads. If a real lab padded a safetensors header the same way to inflate a leaderboard number, nothing in that verification step would catch it.

The entire 8.25-terabyte file traces back to one deduplicated 64-kilobyte block of nulls. The upload moved through Hugging Face's Xet system in about 692 kilobytes, a compression ratio near 11,900,000 to one. It carries an MIT license and a declared context window of 4,294,967,296 tokens.

That's possible because Hugging Face computes a model's parameter count from the safetensors file header, the metadata block listing tensor shapes and dtypes. It never reads the tensor data underneath.

A header can declare any shape it wants, and nothing checks that the bytes match. The mechanism, not the joke, is what Hugging Face would need to fix.