Pirate Face torrents Hugging Face models to keep open weights alive after takedowns

Pirate Face mirrors permissively-licensed open models from Hugging Face — LLMs, image and audio models, and datasets — as BitTorrent magnet links, positioning itself as a decentralized ‘permanence layer’ for open-source AI. Each torrent embeds a BEP-19 web-seed pointing at the model’s file on Hugging Face, so downloads work at full speed with zero peers while the source is live. If Hugging Face removes a model, the web-seed dies and the download falls back to the peer-to-peer swarm, which the project marks as ‘Rescued.’ Integration is meant to be frictionless: setting HF_ENDPOINT=https://pirateface.co reportedly routes an existing pipeline through the service with no code changes, though that endpoint is flagged as not yet live.

The design leans heavily on integrity verification to address the obvious risk of tampered mirrors. Every weight file carries its official Hugging Face SHA-256, so downloaders can confirm the exact bytes regardless of where they pulled them from. Identity is handled separately through optional handle reservation and Hugging Face account verification, which grants a creator badge to prevent impersonation and squatting — the legitimate owner can always reclaim a matching handle. No account is needed to browse, download, or seed. Submissions are currently limited to MIT and Apache-2.0 models (plus a Kimi-K3 exception) that already exist on Hugging Face, with matching LFS hashes at a pinned revision required for automatic listing.

The notable tension is that a project pitching independence from centralized hosting still depends entirely on Hugging Face today — for the source weights, the web-seeds, and identity verification. Pirate Face says direct publishing is planned but not available. The site also dangles a points/leaderboard system and vague future perks like free compute credits and exclusive releases, while explicitly stating there is no token and that points aren’t money. For now it’s best understood as a checksum-verified mirroring and seeding layer that adds censorship resistance on top of Hugging Face rather than a true replacement for it.