Llama.garden Introduces BitTorrent and Web Seeds for Decentralized LLM Distribution
The creator of llama.garden has updated the project to enable fast, decentralized distribution of Large Language Models (LLMs) using BitTorrent and web seeds. Key updates include an API that resolves and caches Hugging Face CDN URLs, open-sourced seed box management scripts, and improved client compatibility. Downloading massive LLM files often strains bandwidth and central servers like Hugging Face. By leveraging peer-to-peer torrents and web seeds, this project provides a scalable, tamper-proof, and faster alternative for the AI community to share and download models. The system uses HTTP-based web seeds for faster initial seeding before relying on peer-to-peer distribution, and the torrent files are cryptographically tied to specific Hugging Face commits to prevent tampering. While compatibility with qBittorrent has improved, the Transmission client is currently recommended for its superior handling of web seeds.
## BACKGROUND
BitTorrent is a decentralized peer-to-peer protocol used for sharing large files, where "web seeds" allow clients to download parts of a torrent directly from traditional HTTP web servers. Hugging Face is the leading repository for hosting open-source AI models, but downloading multi-gigabyte model weights from its Content Delivery Network (CDN) can be slow and resource-intensive. Combining these technologies allows users to download models from both peer-to-peer networks and official servers simultaneously.