How hali works

hali separates discovery, transfer, and attribution into a practical system: web users discover models through a curated registry, teams share efficiently on LAN, and publishers build reputation through signed metadata.

Executive summary

  • hali registry is a curated model catalog, not an open tracker.
  • Magnet downloads work with standard torrent clients.
  • CLI-based sharing is LAN-first, while registry discovery is internet-facing.
  • Publication is telemetry-enabled metadata flow.

Operating modes

Public registry downloads

For anyone downloading from the web

Search on hali.network, copy magnet, open in any BitTorrent client.

Internet-distributed model delivery with Hugging Face webseed support.

Why this exists: Keeps consumption simple: discovery on the registry, transfer on standard torrent tooling.

LAN-first CLI sharing

For teams, labs, and office networks

Use hali CLI to pull from Hugging Face or local peers, then seed locally.

Fast local reuse and lower repeated WAN bandwidth costs.

Why this exists: Turns one upstream fetch into many local downloads without requiring central infrastructure.

Registry publication

For model publishers and maintainers

Download via hali with telemetry enabled to publish metadata to the registry.

Discoverable listings with publisher attribution and trust signals.

Why this exists: Maintains provenance and quality controls while avoiding manual publish workflows.

System flow

Hugging Face originhali metadata + hashingsigned registry listingmagnet distribution + webseed fallback

Technical details

Signed manifests: publisher keys use Ed25519, with canonical JSON and SHA-256 digest verification.

Identity: model identity is canonicalized by infohash (v2 preferred, v1 supported).

Trust scoring: weighted blend of format validity, publisher reputation, download success, and maturity; low-score entries are hidden.

Publication model: registry publication occurs via telemetry-enabled metadata flow.