Back
Pirate Face Turns Open AI Models into Torrents That Survive Deletion
SiTech Team3 წთ. საკითხავი

Pirate Face Turns Open AI Models into Torrents That Survive Deletion

A new peer-to-peer service called Pirate Face mirrors open models from Hugging Face as checksum-verified torrents, promising that weights stay available even if the original host removes them.

Pirate Face, a peer-to-peer service launched at pirateface.co, mirrors open AI models from Hugging Face as torrents, with the stated goal of keeping weights available after any single host takes them down. The project describes itself as a decentralized layer for sovereign AI: models — LLMs, image, audio and datasets — spread as magnet links that, in its wording, "can never be taken down."

Web-seed now, swarm later

Each listing is a magnet link with a Hugging Face web-seed built in — a plain HTTPS URL embedded in the torrent under BitTorrent spec BEP-19. While a model stays on Hugging Face, downloads come straight from the original host at the same speed, and the torrent works even with zero peers. If Hugging Face removes the model, the web-seed dies and downloads fall back to the peer-to-peer swarm; the entry is then marked "Rescued" and remains reachable as long as someone seeds it. Pirate Face argues this removes the single point of failure without changing the bytes users receive.

Checksums against tampered copies

Every weight file carries its official Hugging Face SHA-256 hash, verified at download time — an answer to the common worry that mirrors may serve altered files. Accounts are optional: browsing, downloading and seeding work without one. Anyone may reserve a handle, but the verified badge requires proving control of the matching Hugging Face account, and a reserved name can be reclaimed by its real owner. The FAQ states verification is not needed to use the service.

Drop-in endpoint and submission rules

A drop-in API is promised: setting HF_ENDPOINT to pirateface.co should let existing pipelines resolve models through the service with no code changes, serving from Hugging Face while it is up and from the swarm when it is not. Submitting a model currently requires the weights to be on Hugging Face first, licensed MIT or Apache-2.0 (plus an approved Kimi-K3 exception), with a peer-only magnet, a pinned revision, file checksums and license evidence; listings stay private until review. Accounts earn points — 25 per approved model up to 20 models, 50 per credited referral up to ten — and the site states there is no token.

Model removals from major hubs have repeatedly sent users hunting for copies, and a hash-verified torrent layer offers an auditable alternative. Pirate Face still depends on Hugging Face today and direct publishing is not live, so the promise of permanence rests on how many people choose to seed.

SSiTech

SiTech — AI-powered web development

We build fast, modern websites and bring AI into real business workflows. Have a project or a question? We'd love to help.