Pirate Face mirrors Hugging Face's open models as torrents

Pirate Face has launched as a decentralized layer for open-weight AI models. It mirrors MIT and Apache-2.0 licensed models from Hugging Face, plus one named exception, Kimi-K3, turning each into a checksum-verified BitTorrent magnet link held by a peer-to-peer swarm rather than by a single company or server.

Every mirrored file carries its official Hugging Face SHA-256 checksum, so a download from any source can be confirmed against the original weights. Each magnet link also embeds a Hugging Face "web-seed": a plain HTTPS link to the file on Hugging Face, built directly into the torrent per the BitTorrent web-seed specification, BEP-19. While a model stays on Hugging Face, downloads pull straight from Hugging Face over that web-seed at the same speed, still checksum-verified. If Hugging Face removes the model, the web-seed link dies and the download falls back automatically to the peer-to-peer swarm; Pirate Face marks that model "Rescued," meaning it stays reachable as long as someone keeps seeding it.

The service is meant to work as a drop-in replacement for existing Hugging Face pipelines: setting the HF_ENDPOINT environment variable to pirateface.co is supposed to redirect requests through Pirate Face with zero code changes, though the FAQ marks this "soon" rather than live today. Browsing, downloading and seeding all work without an account. An account lets someone claim a public handle at pirateface.co/yourname and, by verifying a matching Hugging Face identity, earn a creator badge meant to prevent impersonation; only the first eligible handle per account earns welcome points. Points are explicitly not money or a compute-credit balance: a new account earns 50 points for a first referred handle, up to 10 credited referrals per referring account, and 25 points are credited per distinct rescued model to its first verified handle. Pirate Face says it is preparing further community benefits, including free compute credits and exclusive model releases, but eligibility, limits and launch dates are not yet announced, and there is no associated token.

Submitting a new model today requires an existing Hugging Face account, and the model must already be hosted on Hugging Face; direct publishing to Pirate Face without first uploading to Hugging Face is planned but not yet live. New submissions are restricted to MIT and Apache-2.0 licensed models, apart from the Kimi-K3 exception. A submission whose Hugging Face LFS hashes match the pinned revision is listed immediately; anything that does not match, or whose source has disappeared, stays in a review queue.

Key facts

Pirate Face mirrors MIT and Apache-2.0 licensed Hugging Face models, plus one named exception, Kimi-K3, as checksum-verified BitTorrent magnet links.
Each magnet link embeds a Hugging Face web-seed (BitTorrent spec BEP-19): downloads pull straight from Hugging Face while a model is live and fall back automatically to the peer-to-peer swarm, marked "Rescued," once Hugging Face removes it.
A drop-in HF_ENDPOINT redirect would let an existing pipeline resolve models through Pirate Face with zero code changes, but the FAQ marks this feature "soon," not yet live.
Accounts are optional, browsing, downloading and seeding need none, and earn points rather than money: 50 for a referred handle (up to 10 per referring account) and 25 per rescued model; there is no token.
Submitting a new model requires an existing Hugging Face account and that the model already be hosted on Hugging Face; direct publishing without Hugging Face is planned but not yet available.

Why it matters

Open-weight models currently depend on Hugging Face's continued willingness to host them. Pirate Face's pitch is to decouple long-term availability from that single platform by keeping a checksum-verified copy of each permissively licensed model circulating in a peer-to-peer swarm, so removal by the host no longer means the model disappears for everyone who already has access to it.

Who it affects

Developers and teams whose pipelines pull open-weight models from Hugging Face, particularly under permissive licenses like MIT or Apache-2.0; model creators who want to claim a public handle and a verified badge tied to their Hugging Face identity; and anyone who wants to seed or download a model without needing an account at all.

How to use it

Browsing, downloading and seeding require no account. Claiming a handle at pirateface.co/yourname and earning the verified creator badge requires proving a matching Hugging Face identity. Submitting a new model requires an existing Hugging Face account and hosting the model on Hugging Face first; only MIT and Apache-2.0 licensed models are accepted, apart from the Kimi-K3 exception. The planned HF_ENDPOINT drop-in redirect for existing pipelines is marked "soon," not yet live.

How solid is it

The FAQ is detailed about the mechanics it claims: SHA-256 checksum verification against Hugging Face's own hashes, the BEP-19 web-seed mechanism, and specific license and points figures. But the page does not name the company behind it, its founders, funding or a launch date, nor does it state how many models are currently mirrored. Several described features, direct publishing, the drop-in API, and seeding rewards, are explicitly marked planned rather than live, so the working feature set is narrower than the full pitch.

Risks and caveats

The source offers no discussion of the legal or copyright status of re-mirroring licensed models this way; its FAQ addresses only checksum integrity and handle impersonation. Long-term permanence for a "Rescued" model depends on enough people actively seeding it, and Pirate Face itself says seeding rewards are planned but not active, pending attributable seeding evidence. Points do not automatically convert into the promised compute credits, and eligibility, limits and launch dates for those benefits remain unannounced.

“Turn AI into torrents that live forever.”

— Pirate Face