Pirate Face mirrors open Hugging Face models as checksum-verified magnet links held by a peer-to-peer swarm

Pirate Face launched this weekend as a peer-to-peer mirror of open models on Hugging Face, treating each model as a BitTorrent magnet link that survives the original page being taken down. A Show HN post about the site reached 121 points and 41 comments in about two hours on 2026-09-20.

The service currently lists more than 669,000 eligible Apache-2.0 and MIT models, synced live from Hugging Face and re-published as magnet links. Every weight file carries its official Hugging Face SHA-256, so the same hash has to match wherever the file is downloaded from. Pirate Face says this is the point: "the real weights, never a tampered copy."

How the fallback works

Each torrent ships with what BitTorrent calls a web-seed, a plain HTTPS URL to the model file on Hugging Face that is guaranteed to work even with zero peers. While the file is on Hugging Face, downloads pull straight from there at the same speed. The moment Hugging Face removes it, the web-seed dies and the download falls back to whichever machines are seeding the torrent. Pirate Face marks that state as "Rescued" on the model page.

The site also offers a drop-in API. Setting HF_ENDPOINT=https://pirateface.co routes an existing pipeline through Pirate Face instead of huggingface.co, with the same paths and the same API. Pirate Face labels that feature "soon", not live today.

What is on the site now

Trending listings on the home page name specific models with their seed counts and Hugging Face download totals: deepseek-ai/DeepSeek-V4.1-Flash at 765 GB and 7 seeders, Qwen/Qwen3.8-27B at 56 GB and 6 seeders, sentence-transformers/all-MiniLM-L6-v2 at 91 MB and 18 seeders, plus abliterated Qwen and Gemma variants and the recently released convaiinnovations/laya classifier. The activity feed shows verified handle claims by named Hugging Face organisations, including Z.ai for its GLM foundation models.

Anyone can browse, download and seed without an account. Claiming a handle at pirateface.co/yourname is optional, and the "Verified creator" badge requires proving control of a matching Hugging Face account or organisation. Direct publishing on Pirate Face itself, without first putting a model on Hugging Face, is planned but not live: today a submission needs the model to already exist on Hugging Face and match its LFS hashes at a pinned revision.

The scope is Apache-2.0 and MIT

Pirate Face restricts what it lists to Apache-2.0 and MIT models, with a single named exception for Kimi-K3. That covers most of the open weights teams already depend on for inference and fine-tuning, and leaves closed and restrictively licensed models out. The site describes itself as "decentralized infrastructure for sovereign AI" and its short-term goal as making open-model availability outlive any single hosting platform. The FAQ notes that today parts of Pirate Face still depend on Hugging Face and its community.

The reader's action here is small and worth taking early. When a project standardises on an open model, add a Pirate Face URL alongside the Hugging Face one in the dependency lock, and note the checksum. If the day comes that the model disappears from Hugging Face, swapping HF_ENDPOINT restores the same bytes with no code change and no wait for a legal process. The cost is zero while the model is on Hugging Face, and a single takedown, a licence rewrite, or a regional block no longer removes the model from the reachable set.

Source

Pirate Face landing page (Show HN discussion, 2026-09-20)