Pirate Face mirrors Hugging Face models as torrents that survive deletion

How the mirroring works

Pirate Face, a new project that surfaced on Hacker News, mirrors open-weight models from Hugging Face as BitTorrent torrents and keeps them available through a peer-to-peer swarm. According to the project's site, every mirrored model is a magnet link with a built-in web-seed — in BitTorrent terms, a plain HTTPS URL (extension BEP-19) pointing at the model's files on Hugging Face.

That construction makes the fallback automatic. While Hugging Face still hosts a model, a torrent client pulls the bytes straight from the platform at normal speed and the swarm is effectively idle. If Hugging Face removes the model, the web-seed dies and the download falls back to whoever is seeding the torrent; Pirate Face marks such listings "Rescued". The project also claims the swarm can serve regions far from Hugging Face's infrastructure faster than the original host.

Integrity and creator identity

The site says every file carries its official Hugging Face SHA-256 checksum, so a download from any peer has to match the original release bit-for-bit before it counts. That check targets the main fear with mirrored weights: ending up with a tampered copy.

Identity is handled separately. Anyone can browse, download and seed without an account, and reserving a handle proves nothing. A "Verified creator" badge requires proving control of the matching Hugging Face account, and the real owner of a reserved name can reclaim it by verifying, which the project says stops squatters from locking creators out.

What is live and what is promised

For now the service leans on Hugging Face existing. A submission requires a Hugging Face account, a model already hosted there, a magnet, a pinned revision, file checksums and license evidence. Only MIT and Apache-2.0 licenses are accepted, along with one named exception for Kimi-K3. Listings whose checksums match Hugging Face's LFS hashes at the pinned revision go live immediately; mismatches or dead sources sit in a review queue.

Several features are announced but not live yet. A drop-in API would let existing pipelines point HF_ENDPOINT at Pirate Face and keep working with zero code changes, resolving from Hugging Face while it is up and from the swarm once it is not. Direct publishing, meaning uploads that skip Hugging Face entirely, is also planned. There is a points scheme for participation — 50 points per qualifying referral, 25 per rescued model — with seeding rewards still to come, and the site states plainly that there is no token.

Why it matters

Open-weight models are usually open only in licensing terms; in practice, most people download them from a single company. If that host deletes a model, gates it behind approval, or is compelled to take it down, the weights become hard to get no matter what the license says. Pirate Face replaces that single point of failure with a swarm, and the web-seed design means the safety net costs nothing while the original is alive.

The caveats are real. The project currently depends on Hugging Face for checksums, verification and initial listings, so it is not independent of the platform it hedges against. Its catalog only works if people actually seed, and redistributing weights peer-to-peer can raise licensing questions even for permissive licenses. As an early signal, though, it sketches a plausible shape for open-source AI infrastructure: weights verified by hash and hosted by nobody in particular.