Skip to main content

hali

Download once. Seed for everyone.

hali is the fastest way to grab an LLM model and share it with your team, your LAN, and — if you choose — the entire world.

It's a CLI tool. It's a local cache. It's a background daemon that turns every download into a BitTorrent seed. No accounts. No cloud sync. No "enterprise platform." Just a single binary that does one thing: gets models onto your disk and makes them available to everyone else.

The name? hali is Finnish for hug. As in Hugging Face. You see what we did there.

$ hali pull mistral

That's it. One command. hali searches Hugging Face for the best GGUF match, downloads it via HTTP, hashes the pieces in the background, turns it into a torrent, and starts seeding — all automatically. You didn't ask for any of that. It just happened.

Why you'll like it​

It's one command​

hali pull mistral. Three words. hali figures out the repo, the quantization, the file. You answer one prompt (or skip it with a model ID like mistral:7b:instruct:q4_k_m) and the download starts. No config file needed. No API keys. No "create an account."

LAN acceleration, automatically​

The moment you download a model, hali starts broadcasting its availability to every machine on your local network via UDP multicast. Anyone else on your LAN running hali will see it and download from you at local speed instead of hitting Hugging Face again. A 70 GB model downloaded once. Everyone else gets it at gigabit speed.

It exports to Ollama and LM Studio​

hali is a distribution tool, not an inference engine. But it talks to the tools that do inference. One command to create an Ollama manifest (instant — no file copy). Another to drop the GGUF into LM Studio's models directory. One more to do both at once.

It runs as a service​

Install it as a Windows service or a Linux systemd unit. The daemon starts on boot, seeds your models in the background, and keeps everything running. System tray app on Windows. journalctl on Linux. You'll forget it's even there.

It turns you into a philanthropist​

Here's the thing that makes hali different: every download is a contribution to the network. When you pull a model, hali automatically generates a torrent from it and submits it to the public registry. That torrent — signed with your Ed25519 key — becomes a verified entry anyone can find. Downloads always work because the torrent embeds a webseed (the HuggingFace CDN URL), so even if nobody is seeding yet, the first person who discovers your entry can grab the model directly from HuggingFace. Once they do, they become a seeder. The network grows with every pull.

You didn't have to do anything extra. You just wanted a model. And now countless other LLM enthusiasts get faster downloads. You're basically a digital philanthropist. Put that on your resume.

The whole picture​

hali is part of a three-piece ecosystem:

Three independent pieces. The CLI works offline. The daemon works without internet. The registry is optional. Each piece can fail and the others keep running.

Ready?​

Pick your path:

  • Quick Start — get a model in under 60 seconds
  • Windows Setup — install, configure the service, set up the tray app
  • Linux Setup — .deb package, systemd, permission groups
  • Why hali? — the problem we're solving and why it matters
  • How It Works — deep dive into the architecture, trust scoring, and torrent lifecycle