How to seed & run a mirror

Seeding is how the community keeps models available. Here's how to help.

Install the CLI (recommended)

If you already have model files on disk, the llmt command-line client is the fastest way to seed them. It scans the model files you already hold, matches each file's SHA-256 against the catalog, then hands the confirmed ones to your torrent client to seed — without re-downloading the model weights (small companion files such as a license or README may still be fetched).

Install it with pip (Python 3.9+); this installs the llmt command:

pip install llmt-cli        # or, for an isolated install:  pipx install llmt-cli

Then point it at the folders where your models live. llmt scan finds locally-held model files and matches each file's SHA-256 against the catalog (a catalog lookup, not an Ed25519 signed-manifest check — matches are reported as catalog-matched, never verified), builds a small link farm (hard- or symlinks — nothing is copied or moved), and hands the matches to qBittorrent or Transmission so you seed what you already have:

llmt scan           # match local models to the catalog, then seed them
llmt doctor         # diagnose why your seeds aren't actually seeding

llmt scan hands matches to your client automatically only if the client's remote control is on: enable qBittorrent's Web UI (Options → Web UI) or Transmission's RPC, then pass its address and credentials with --qb-url/--qb-user/--qb-pass (or the LLMT_QB_URL/LLMT_QB_USER/LLMT_QB_PASS env vars; Transmission has --tr-* equivalents). Without them scan just prints the magnet for you to add by hand. Either way, the client has to stay running for the model to keep seeding.

Integrity. Every release ships a SHA256SUMS file that is minisign-signed; the checksums and the signing public key are published on this site, so you can verify the download before you trust it.

Running qBittorrent in a container? The link farm defaults to ~/.llmt/seeding-layouts. If your torrent client runs in Docker, point the farm at a path the container can see with llmt scan --layout-base <path> so the links resolve inside the container. Container-path ergonomics are being improved in the CLI.

Prefer to do it by hand, or seeding just one model? The manual qBittorrent path below works too.

Seed a model in three steps

  1. Open the model you want to seed on the catalog, accept the license, and add its magnet to a BitTorrent client (qBittorrent, Transmission, Deluge, or similar).
  2. Let it finish downloading. Once complete, your client automatically switches to seeding — uploading pieces to others.
  3. Keep the client open. The longer you seed, the healthier the swarm. There's no ratio requirement and nothing to earn — seeding is voluntary and appreciated.

Pledge to seed

Verified members can publicly pledge to seed a torrent. A pledge is a statement of intent — it tells others you're helping. Because our tracker is anonymous, we can't (and don't) measure exactly who is seeding what, so a pledge is never presented as a verified measurement. It's a thank-you and a signal, nothing more.

Run a mirror

A mirror is a machine (often a seedbox or VPS) that seeds many torrents around the clock. To run one:

  • Provision a box with enough disk for the models you want to host.
  • Run a headless client (qBittorrent-nox, Transmission-daemon) as a service.
  • Add the magnets for the models you want to mirror and leave it running.
  • Open the client's listening port so peers can reach you.

Webseeds

Every torrent here also carries a webseed (an HTTP mirror of the files). Webseeds let a client download directly over HTTPS when peers are scarce, so a model stays available even at very low seeder counts. You don't need to configure anything — modern clients use webseeds automatically. Running your own webseed (a public HTTPS mirror of the files) is an advanced way to help; the webseed URL is embedded in the torrent metadata.

Why it matters

Models with few seeders download slowly and risk disappearing. The needs-seeders queue shows which torrents most need help right now. Even seeding one model for a few days makes a real difference.