llama.garden: A Nostr-Powered Decentralized LLM Torrent Network

llama.garden is a new attempt to distribute large language models over BitTorrent. It tackles the classic problem of seedless torrents by falling back on Hugging Face's own web servers, so downloads work from day one even with zero peers seeding. Right now this fallback only works reliably in Transmission, since qBittorrent fails to follow Hugging Face's redirect chain correctly.

Under the hood, the project uses the Nostr protocol as a decentralized catalog instead of a single website. Model listings are signed events (kind 30099) propagated across nine independently run relays, so no single operator can delete a listing network-wide — only opt out of serving it locally. The client itself is a single, forkable HTML file, and its update mechanism is tamper-evident: new versions are only accepted if their SHA-256 hash matches one signed by the curator, preventing a compromised host from pushing malicious code.

Recent updates include showing seeder counts on model cards, faster HTTP web-seeding via nginx configured for both TCP and UDP traffic, community-requested magnet link support (though .torrent files remain more reliable), and a submission/approval workflow for community-contributed models.

For engineers, the project demonstrates a practical, censorship-resistant, cryptographically verifiable way to distribute large model weights without relying on centralized hosting or a single point of failure.