Skip to main content

Why hali?

Every AI lab, homelab, and enthusiast with more than one machine has the same problem: you download a 70 GB model, and then you download it again. And again. And again.

Your workstation needs it. The build server needs it. Your teammate's laptop needs it. That Raspberry Pi cluster you're "experimenting with" needs it. Every single one of these machines reaches out to Hugging Face's CDN and yanks down the same 70 gigabytes, burning bandwidth, waiting in queues, and wasting time.

hali fixes this.

The problem, spelled out​

Four machines. Four downloads. 280 GB of redundant internet traffic. Days of waiting. Same bits, same bytes, transferred four times over the public internet.

This isn't just slow — it's wasteful. You have a gigabit LAN sitting right there, doing nothing, while every machine independently hits the internet for the exact same file.

The hali way​

One download from the internet. Three from the LAN. Total time: however long it takes to download 70 GB once, plus a few minutes to transfer across the local network at wire speed. All of this happens automatically — no one had to configure anything, share a network drive, or remember an IP address.

What makes hali different​

It's not a model manager​

Tools like Ollama and LM Studio are runtime environments — they run models. hali is a distribution layer — it gets models onto disk and shares them. It works with your runtime, not instead of it. Export to Ollama with one command and the model appears instantly, no re-download.

It's not a file sync service​

Dropbox and Syncthing sync arbitrary files. hali is purpose-built for LLM models: it understands GGUF, Hugging Face repos, quantization naming, and the specific structure of model files. It generates BitTorrent metadata deterministically so that two people who independently download the same model version produce the same torrent — making their pieces interchangeable.

It's not just for your LAN​

The public registry at hali.network means your torrents become discoverable worldwide. This isn't a tracker for Linux ISOs — it's a curated catalog of verified LLM models with signed manifests, trust scoring, and provenance tracking.

The philanthropy angle​

Here's the part that makes hali genuinely different from every other model downloader:

Every time you hali pull, you're making the internet better for everyone else.

When you download a model, hali automatically:

  1. Generates a BitTorrent infohash from the file
  2. Creates a signed manifest with your Ed25519 key — proving you published this exact model
  3. Submits the torrent metadata to the public registry
  4. Starts seeding on your LAN — other machines on your local network can grab it at wire speed

You didn't opt in. You didn't click "share." It just happened, automatically, in the background, while you were already downloading the model for yourself.

The result: every download grows the network. Someone in São Paulo discovers your model on hali.network and downloads it via the embedded webseed (HuggingFace CDN) — the torrent always works, even with zero seeders. Now they have the file too. They'll produce the same infohash because hali's torrent generation is deterministic. Their pieces are interchangeable with yours. The registry gets broader coverage, the torrent gets more potential peers, and everyone wins — all without you lifting a finger.

You are now a digital philanthropist. You downloaded a model for your own use. In doing so, you automatically created a verified torrent, published it to the global registry, and started seeding it on your LAN. Anyone in the world can now discover it on hali.network and download it via the embedded webseed — no seeders required. And once they download it, they become seeders too. The swarm grows with every pull. No extra clicks. No "share" button. It just happened.

The design philosophy​

hali follows a few hard rules:

  • Local-first. Everything works without internet. The cache is on your disk. The daemon runs locally. LAN sharing needs no cloud.
  • Zero configuration by default. You can tune settings, but you don't have to. The defaults are sane. The config file materializes itself on first run.
  • Fail gracefully. If Hugging Face is down, the LAN might save you. If the LAN is empty, Hugging Face will do. If the daemon crashes, the CLI carries on. No single point of failure.
  • Privacy-respecting. Telemetry is opt-in. The daemon binds to 127.0.0.1 by default. LAN announcements can be HMAC-authenticated. Nothing phones home without your permission.
  • Deterministic identity. Two people downloading the same model version produce the same torrent infohash. This is the property that makes cross-seeding work. It's not accidental — it's enforced by the spec.

Want the full technical story?​

Head to How It Works for the deep dive: the pull lifecycle, torrent generation, LAN protocol, trust scoring, and the Ed25519 signing pipeline.

Ready to get started? Jump to the Quick Start.