16.5T Vacuum Model Exploits Hugging Face Headers to Top Leaderboards
alerikaisattera · reddit · 2026-08-02
A developer exploited Hugging Face's parameter counting mechanism—which sums prod(shape) from safetensors headers without reading tensor data—to top the Hub's leaderboard with a completely empty 16.5-trillion-parameter model named Vacuum-16T.
While all weight bytes are strictly 0x00, the project reveals several concrete technical findings about model hosting:
- Bandwidth vs. Storage Quota: Thanks to Xet content-defined chunking deduplication, the actual network transfer was only 692 KB (an 11.9M:1 ratio). However, storage quotas do not deduplicate, meaning the repo still logically consumes its full declared 8.25 TB.
- Metadata Cost: The only irreducible cost in an empty model is naming. Using 65536×65536 tensors requires 4,000× less metadata overhead than 1024×1024 tensors for the exact same declared parameter count.
- Extreme Context Window: The author declared a maxpositionembeddings of 232 (4.29 billion tokens), backed by a real 2.15 GB position-embedding tensor of zeros to pass the parser.
Billed as the "safest AI model" that refuses 100/100 jailbreaks, the repo is a technical troll against labs obsessing over raw parameter counts.
More from Fun
- Meme: Cybertruck lifted by rocket thrusters celebrates 'AI that does all your work' — coherence · 2026-10-16
- Tokyo's Gemini AI Photo Booth turns visitors into plush toys and dioramas — bradneuberg · 2026-10-16
- AI face-swap of Seinfeld's George Costanza is uncannily convincing — alanskimp · 2026-10-16
- Tinyfool sparks debate over whether Tesla Roadster 2 counts as 'useless' innovation — tinyfool · 2026-10-16
- Beff Jezos: when a cracked autist starts detailing their plans, believe them — beffjezos · 2026-10-16
- Tinyfool: does Xiaomi's foldable released days before iPhone count as innovation? — tinyfool · 2026-10-16