Skip to main content

Quick Start

Get a model onto your disk and into your runtime in under 60 seconds.

Prerequisites​

  • A machine running Linux, macOS, or Windows (x86_64 or ARM64)
  • A terminal
  • That's it

No API keys. No accounts. No config files. The binary is self-contained.

Step 1: Install hali​

Linux​

# .deb package (recommended)
sudo dpkg -i hali_0.1.0_amd64.deb

# Or manual tarball
tar xzf hali-linux-amd64.tar.gz
sudo cp hali halid /usr/bin/

macOS​

# Homebrew (coming soon)
# brew install hali

# Or manual download
curl -fsSL https://hali.network/install.sh | sh

Windows​

# Download the .msi installer
# Or extract hali.exe and halid.exe to a directory in your PATH

hali --version
# hali version 0.1.0 (abcdef1) oss

Full installation details: Windows Setup | Linux Setup

Step 2: Search for a model​

hali search mistral

Output:

  1  TheBloke/Mistral-7B-Instruct-v0.2-GGUF       2.1M downloads
  2  bartowski/Mistral-7B-Instruct-v0.3-GGUF       840.5K downloads
  3  MaziyarPanahi/Mistral-7B-Instruct-v0.3-GGUF   320.1K downloads
  ...

Use 'hali pull <repo>' to download a model.

hali queries HuggingFace for GGUF models, ranked by download count. No fluff — just the repos.

Step 3: Pull a model​

hali pull mistral

If you give a bare search term, hali goes interactive:

  1. Pick a repo from the ranked list
  2. Pick a quantization — sorted smallest to largest
Choose a repo (1-3): 1
Selected: TheBloke/Mistral-7B-Instruct-v0.2-GGUF

Available files:
  1  mistral-7b-instruct-v0.2.Q2_K.gguf        2.83 GB
  2  mistral-7b-instruct-v0.2.Q3_K_M.gguf      3.26 GB
  3  mistral-7b-instruct-v0.2.Q4_K_M.gguf      4.14 GB
  ...

Choose a file (1-10): 3

The download starts. You'll see a progress bar:

[████████████████░░░░░░░░] 2.1 GB / 4.1 GB  51%  18.2 MB/s

When it finishes:

Saved mistral:7b:instruct:q4_k_m (4.14 GB)
  magnet: magnet:?xt=urn:btih:a3f9c21b4d67...

Or skip the prompts entirely​

# Use a direct HuggingFace repo path
hali pull TheBloke/Mistral-7B-Instruct-v0.2-GGUF

# Use the canonical model ID (fastest)
hali pull mistral:7b:instruct:q4_k_m

Step 4: What just happened automatically?​

While you were watching the progress bar, hali did several things without asking:

  1. Hashed the file into 16 MiB pieces while downloading (if streaming_hash is on)
  2. Computed a torrent infohash from those pieces
  3. Built a .torrent file and a magnet URI
  4. Launched the daemon (if it wasn't already running)
  5. Started seeding the model via BitTorrent on your LAN
  6. Submitted a signed manifest to hali.network — your model is now globally discoverable

You just became a philanthropist. People on your LAN can get this model from you at local speed. People on the internet can now discover it on hali.network and download it via the embedded webseed — and once they do, they become seeders too.

You just made the network better for everyone. Your download created a verified torrent entry in the public registry. Anyone searching for this model on hali.network can now discover it and download it via the embedded webseed (HuggingFace CDN). When they do, they become seeders — and the swarm grows stronger with every pull. Zero extra effort on your part.

Step 5: Check your cache​

hali list
MODEL ID                                    SIZE        DOWNLOADED
------------------------------------------  ----------  ----------
mistral:7b:instruct:q4_k_m                 4.14 GB     2026-05-23

Step 6: Check the daemon​

hali daemon status
Daemon running  PID 12345  uptime 5m22s  port 51234

SEEDING                                     STATUS    PEERS
------------------------------------------  --------  -----
mistral:7b:instruct:q4_k_m                 seeding   0 peers
  magnet: magnet:?xt=urn:btih:a3f9c21b4d67...

The SEEDING section shows what's being shared. The magnet link is the same one someone can paste into any BitTorrent client to download this model from you.

Step 7: Use the model in a runtime​

hali is a distribution tool. To actually run the model, export it to a runtime:

# See what runtimes are installed
hali runtime list

# Export to Ollama (instant — creates a manifest, no file copy)
hali export ollama mistral:7b:instruct:q4_k_m

# Export to LM Studio (copies the GGUF into its models dir)
hali export lmstudio mistral:7b:instruct:q4_k_m

# Do both at once
hali export all mistral:7b:instruct:q4_k_m

After hali export ollama, run ollama list — the model appears immediately without re-downloading.

Step 8: Open the web dashboard​

hali stats --web

Opens http://127.0.0.1:47433 — a live dashboard showing:

  • Download and upload speeds
  • Session transfer totals
  • Active model rows with status and peer counts
  • Clickable magnet links

Step 9: Keep it running across reboots (optional)​

The daemon auto-starts when you hali pull. But it stops when you close your terminal. To keep it running permanently:

hali service install   # Register with SCM (Windows) or systemd (Linux)
hali service start     # Start the service
hali service status    # Confirm it's running

Now the daemon starts on boot, restarts on crash, and seeds your models 24/7. LAN peers can always find you.

What's next?​

You just did this​

Three words in a terminal. Nine steps handled automatically. You now have a model on disk, a torrent in the registry, and a daemon seeding to your LAN. All without creating an account, configuring anything, or clicking a "share" button.