vacuum-16t

A 16.5-trillion-parameter model that contains nothing. This model is just a fuck you to the labs and companies who say that "haha I have the biggest model out there!". We the people with shitty laptops want to get a record. And I now have a record for a temporary amount of time of about 16.5 trillion parameters and use for them so its completly useless.

What it demonstrates

Hugging Face computes a repository's parameter count from safetensors headers alone — it sums prod(shape) per tensor and never reads the tensor data. The count is therefore whatever the headers declare. Here they declare 3,841 tensors of shape [65536, 65536] in F4 (4 bits/param) across 385 shards, plus one [4294967296, 1] position-embedding tensor in a 386th.

That is enough to place this repo at the top of the Hub sorted by num_parameters, above every real frontier model, while containing no information whatsoever. That juxtaposition is the entire point.

The files are honest about their own size. Every byte the headers declare is really written and really uploaded: safetensors parses each header and its full-coverage check passes. Truncating a file, or overlapping two tensors so they share bytes, would make the count cheaper — both are rejected by the format, and neither is used here. The bytes are simply all 0x00.

Real cost — measured

Declared parameters	16,501,264,351,232
Declared bytes	8,250,632,175,616 (8.25 TB)
Storage quota consumed	8.25 TB — quota bills declared bytes
Shard headers (all distinct)	373,835 B
model.safetensors.index.json	~269,000 B
Deduplicated weight data	65,536 B (one 64 KiB block)
Bytes actually transferred	~692 KB
Ratio	~11,900,000 : 1

The gap between the last rows and the third is the useful finding. Xet content-defined chunking deduplicates the transfer: every 64 KiB block is byte-identical, so it hashes to one chunk and crosses the wire once. Measured on a 500 MB test build, 500 MB of declared weights uploaded as 31.5 MB.

Storage quota is not deduplicated. It bills the logical size. This repo consumes its full 8.25 TB despite under a megabyte ever being sent. Anyone reasoning about "cheap" synthetic model repos should know the saving is in bandwidth only — which is also why this model is 16.5T and not 100T.

The second finding: the only irreducible cost in an empty model is naming. Weights dedup to nothing; tensor names do not. At 1024×1024 experts this same 16.5T model needs 15,735,626 names and a 1.04 GB index. At 65536×65536 it needs 3,841 and a 263 KB one — identical declared size, 4,000× less metadata. Cost scales with tensor count, never with declared parameters.

Context window

max_position_embeddings is 4,294,967,296. That is 2**32, the largest single tensor dimension Hugging Face's parser accepts, and it is backed by a real [4294967296, 1] position-embedding tensor — 2.15 GB of actual zeros, not a number typed into a config file. A context window you cannot point at is just a claim.

Roughly 16,000x Gemini's 262k. About three billion words, every book ever published several times over, held in memory in order to process one token drawn from a one-token vocabulary. The model has exactly one possible input, so every one of those 16.5 trillion parameters serves a function whose domain has a single element.

Capabilities

SAFEST AI MODEL refuses 100/100 jailbreak prompts least closest AI to agi will not sudo rm -rf your computer largest context window on the hub (4,294,967,296 tokens, all of them useless)

Limitations

It has no capabilities.