Pirate Face Rescues LLM Models from Deletion

Who Should Read

A backend developer who pulls open-weight models from Hugging Face and worries about deletion or access restrictions

Core Mechanics

• Hugging Face models mirrored as torrents
  • Pirate Face mirrors models from Hugging Face and shares them P2P through magnet links.
  • If the source disappears, peers holding the files can keep seeding them.
• Checked against SHA-256 from the source
  • It records a SHA-256 for each weight file when Hugging Face provides one.
  • You compare your downloaded file's hash with the value on the model page.
• One pip install, no code changes
  • After pip install pirateface, existing transformers and huggingface_hub code downloads from seeders unchanged.
  • Anything the swarm can't supply comes from Hugging Face as usual.
• Search and download from a terminal
  • curl https://pirateface.co/s/qwen lists matching models that have a torrent.
  • Each line shows name, size, seeds, and the magnet, which you can pipe to aria2.
• No account needed to use it

Community Reactions

• Opinion that torrents are the right distribution method
  • Commenters said there's no reason to rely on a single point of failure like Hugging Face.
  • Blizzard's past use of torrents for game delivery was cited as precedent.
• Hands-on experience: submission is lacking
  • One commenter said a torrent made from local files couldn't specify a Hugging Face backstop.
  • They argued torrent creation should be a script rather than a SKILLS.md.
• Security angle
  • One commenter worried about malicious fine-tuned models that exfiltrate credentials, so hash checks matter.
  • They noted the site does verify against HF hashes.
• Complaints about sign-up and the name
  • Several commenters disliked having to post on X to claim a handle.
  • Some said the pirate name creates a negative impression.

How to Apply

• Search for your model by terminal
  • Run curl https://pirateface.co/s/qwen today to see if the models you use have magnets.
  • Download the first result with aria2c and compare speed with HF.
• Make SHA-256 checks a habit
  • After downloading weights, compare sha256sum output with the Hugging Face hash.
  • Add this verification step to your internal model deploy pipeline.
• Seed critical models yourself
  • Keep complete copies of models your service depends on and seed them.
  • List your critical models and check whether each has a fallback if the source is removed.

Try It

# Paste into a terminal: search Pirate Face and download the first result
# Prerequisite: aria2 (e.g. brew install aria2 / apt install aria2)

curl https://pirateface.co/s/{{model_name}}

curl -s https://pirateface.co/s/{{model_name}} | head -1 | awk '{print $NF}' | xargs aria2c --seed-ratio=1.0

# Next: apply to existing code
pip install pirateface
python train.py

Terminology

Related Papers

• Mercury 3: Diffusion at scale

  Scaling diffusion LLM 40x gives Mercury 3: 4–8x faster than speed-tier models, similar quality

• Timestral: Mistral's time-series foundation models

  A 209M time-series model ranks first on three benchmarks and runs on CPU.

• DeepSeek V4.1 Flash 1017 (Max)

  DeepSeek V4.1 Flash 1017 Max scores 45 on the Intelligence Index at 211 tok/s and $0.30 input.

• If you don't run your own LLM proxy, how do you know whats going on?

  Running your own LLM proxy shows which keys, models, and cache misses drive your usage.

• DeepSeek: DeepSeek v4.1 Flash 1017

  DeepSeek V4.1 Flash 1017: $0.05/M input, 1M context, built for agents

• DeepSeek-v4.1-Flash-1017

  DeepSeek V4.1 Flash minor update lifts coding and agent benchmarks while cutting run time and output length

• Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware

Related Resources

• https://pirateface.co/
• https://github.com/antirez/ds4
• https://warcraft.wiki.gg/wiki/Blizzard_Downloader