Pirate Face adds a BitTorrent tracker to keep open AI models online
The service pairs Hugging Face web-seeds with peer discovery, but its 669,000-model catalog is far larger than its set of witnessed copies.
By Ryan Merket · Published
Primary source: Pirate Face
Why it matters
Open-weight AI depends on centralized hubs whose owners can gate or remove repositories. Pirate Face tests whether BitTorrent can provide a durable fallback, while exposing the gap between cataloging a model and keeping a complete copy online.
Pirate Face launched a first-party BitTorrent tracker and distributed hash table discovery on September 20th, giving its open AI model torrents two ways to find peers after their original downloads disappear.
The release puts working discovery infrastructure behind Pirate Face's pitch: preserve openly licensed language, image, audio and other machine-learning models by turning their files into checksum-verified torrents. Pirate Face first described that architecture in a July 9th mission essay. The September 20th infrastructure update adds a tracker operated by Pirate Face alongside BitTorrent's distributed lookup network.
Every magnet served by Pirate Face now includes udp://tracker.pirateface.co:6969/announce. A compatible client announces the torrent's infohash to that endpoint and receives addresses for other clients carrying the same files. DHT offers a second route that can continue finding peers without the Pirate Face tracker.
Pirate Face does not relay the model weights through the tracker. Peers exchange the files directly, and the tracker coordinates discovery. That distinction matters because Pirate Face can disappear without necessarily taking an established DHT swarm with it.
Hugging Face remains the primary source
Pirate Face currently builds around Hugging Face rather than replacing it. Each torrent can contain a BEP-19 web-seed pointing to the corresponding Hugging Face download. While the repository remains available, users can retrieve the files from Hugging Face. If that source stops responding, the same magnet can fall back to peers with complete copies.
Pirate Face compares the downloaded files with SHA-256 checksums recorded from Hugging Face. The check establishes that the bytes match a previously witnessed revision. It does not establish that the model is safe, accurate or lawful to redistribute, limitations Pirate Face spells out in its terms of use.
Pirate Face says its catalog covers more than 669,000 models eligible under its Apache-2.0 and MIT licensing policy. The larger number describes repositories that could be preserved, rather than hundreds of thousands of working backups.
Pirate Face's activity dashboard reported 4,891 witnessed records on September 20th. A witnessed record means Pirate Face saved revision and checksum evidence. It does not mean Pirate Face stored the weights, created a torrent or found an active seeder.
That gap is visible on individual model pages, which separate source metadata, checksum records, torrent registration and peer availability. Pirate Face repeatedly warns that a listed magnet does not prove a complete download is currently reachable.
The public leaderboard showed 18 seeders for its most-seeded model, the 91 MB sentence-transformers/all-MiniLM-L6-v2, at the time of review. Larger entries had thinner replication: Pirate Face reported 11 seeders for a 765 GB DeepSeek model and 11 for a 649 GB Qwen derivative. Those counts are observations from the tracker and DHT, and can change whenever clients disconnect.
That is the operational constraint behind the permanence claim. A magnet link and checksum record can survive indefinitely as metadata. The model survives only when at least one peer retains every file and remains willing to upload it.
A backup layer with takedown limits
Pirate Face accepts community submissions from users holding complete, redistributable copies. Submitters provide a pinned revision, file checksums, license evidence and a peer-only magnet. Pirate Face strips contributor-supplied trackers and web-seeds before adding its own tracker to approved listings.
The system also creates a tension familiar from every peer-to-peer network. Pirate Face's takedown policy says Pirate Face can delist a model, remove its magnet and stop seeding from infrastructure under its control. Pirate Face cannot recall copies already held by independent peers.
Pirate Face limits automated mirroring to models carrying approved licenses, although licensing metadata and the right to redistribute a particular artifact can still be disputed. Pirate Face places that responsibility on seeders and submitters.
The network also inherits BitTorrent's privacy trade-off. Pirate Face's privacy policy warns that torrent clients connect directly to peers and may expose a user's IP address. Organizations seeking resilient model storage may prefer controlled internal mirrors for that reason, even if they use Pirate Face's checksum and provenance records as a reference.
Pirate Face is building toward closer integration with existing machine-learning workflows. Its site advertises a planned endpoint that would let developers set HF_ENDPOINT=https://pirateface.co and keep using Hugging Face-style repository paths. Direct publishing, account benefits and seeding rewards remain planned features rather than live parts of the preservation system.
The tracker is the first piece that makes Pirate Face's survival mechanism less dependent on a static catalog. It gives seeders a shared rendezvous point and leaves DHT available as the fallback. The remaining work belongs to the swarm: storing complete models that can run from megabytes into terabytes, and keeping them available long after the announcement traffic moves elsewhere.