Somebody Is Torrenting Hugging Face Before Anything Gets Deleted
Pirate Face turns open-weight models on Hugging Face into checksum-verified torrents so that they survive the day the original page goes away. It went up at https://pirateface.co/ and took 367 points on Hacker News, and it currently lists 669,000 plus eligible models.
The design is the interesting part. Every model gets a torrent whose web seed points back at Hugging Face, so while the upstream file exists you are effectively downloading from Hugging Face at Hugging Face speed and quietly seeding to everyone else. When the upstream file disappears, the web seed fails over and the swarm keeps serving, and the model gets flagged Rescued. No flag day, no migration, no moment where anybody has to decide to archive something.
This matters more than it did a year ago. Weights get pulled for license disputes, for safety retractions, for takedown requests, and for the most common reason of all, which is that a lab reorganized its org and nobody was paying the storage bill for an old checkpoint. Reproducibility in this field currently rests on a single company's retention policy, and every paper that says we evaluated against model X is a dead link waiting to happen.
The operators are anonymous. There is an X account and a promise to allow direct publishing soon, and no named team, which is a real risk to weigh if you are thinking about depending on this. The counterargument is that a checksum-verified torrent does not require you to trust the operator at all, only the hash, which is the whole point of doing it this way.
The name is a provocation and the function is a library. Those tend to be the same thing at the start.
← Back to all articles
The design is the interesting part. Every model gets a torrent whose web seed points back at Hugging Face, so while the upstream file exists you are effectively downloading from Hugging Face at Hugging Face speed and quietly seeding to everyone else. When the upstream file disappears, the web seed fails over and the swarm keeps serving, and the model gets flagged Rescued. No flag day, no migration, no moment where anybody has to decide to archive something.
This matters more than it did a year ago. Weights get pulled for license disputes, for safety retractions, for takedown requests, and for the most common reason of all, which is that a lab reorganized its org and nobody was paying the storage bill for an old checkpoint. Reproducibility in this field currently rests on a single company's retention policy, and every paper that says we evaluated against model X is a dead link waiting to happen.
The operators are anonymous. There is an X account and a promise to allow direct publishing soon, and no named team, which is a real risk to weigh if you are thinking about depending on this. The counterargument is that a checksum-verified torrent does not require you to trust the operator at all, only the hash, which is the whole point of doing it this way.
The name is a provocation and the function is a library. Those tend to be the same thing at the start.
Comments