Preview
CLI download
uvx -p 3.13 --from pirateface pf download org/model
  • Check the model files with Hugging Face
  • Look for peers in the Pirate Face swarm
  • Use Hugging Face if the swarm cannot help
  • Save verified files in the Hugging Face cache

Your CI pulls a 15 GB model, the download from Hugging Face stalls halfway, and the job goes red. Pirate Face lists 37,157 models as torrents, so the Pirate Face CLI can get the same files from people who already have them, and goes to Hugging Face only for what the swarm can't send.

Pirate Face CLI is the Swiss Army knife of model distribution via torrents.

Your code stays the same

Already using transformers or huggingface_hub? Pull from Hugging Face just like before. If it's ever down, the swarm takes over.

Your scriptvia piratefaceHugging FacefirstPirate Face swarmbackupHugging FacebackupPirate Face swarmfirst

The default: each file comes from the swarm, and Hugging Face steps in when the swarm can't help.

With uv installed, add pirateface to a single run:

PF_HF_FIRST=1 uv run --with pirateface python your_script.py

Or install it once with pip and set PF_HF_FIRST=1 in your environment for the same:

pip install pirateface

Downloads work on their own too. Like hf download, this puts a model in your Hugging Face cache:

uvx -p 3.13 --from pirateface pf download org/model

Swap org/model for any Hugging Face model ID.

Beyond downloads

pf can seed what you already have. pf seed --hf-cache goes through your Hugging Face cache one model at a time and asks before seeding each one. Anything Pirate Face doesn't list yet gets submitted along the way.

uvx -p 3.13 --from pirateface@latest pf seed --hf-cache
+pts+pts+ptsYour diskHugging Face cacheOther peopledownload from youPirate Facepoints for every hour seedednew models get listed once

Short on disk? pf work takes the next model from the rescue queue, streams it from Hugging Face without writing a byte to disk, then builds the torrent and submits it.

uvx -p 3.13 --from pirateface@latest pf work
Rescue queueone model at a timeHugging Facestreams the filespf workbuilds it in memoryPirate Facelisted, earns points

Your own models can go on the list too (pf submit --direct), and the torrents you already seed get kept in shape.

Seeding and rescue work both earn points on the leaderboard.

Even if we go down too

Pirate Face can go down as well. Keep a copy of the magnet index on your machine, and when pirateface.co doesn't answer, pf download looks the model up there and joins its swarm directly. Every piece is checked against the torrent's hashes. Run it once while the site is up, then again now and then to pick up new models.

uvx -p 3.13 --from pirateface@latest pf index --sync
Model swarmstill deliverspirateface.coofflineYour machinewith its index copy

With the site offline, your copy of the index still finds the model, and the swarm still delivers it.

If you'd rather not remember any of this, let your coding agent read skill.md and run it for you.

Source on GitHub