Model Registry uses torrents with Hugging Face fallback for open models

Model Registry (modelregistry.io) is a new open-source project that distributes popular open AI models via BitTorrent, with Hugging Face acting as a web seed fallback. Created by developer Ravindra-Marella and hosted on GitHub (github.com/marella/modelregistry), the system publishes .torrent files that point to model files. When no BitTorrent peers are available, downloads seamlessly fall back to Hugging Face's CDN using the BEP_0019 web seed specification. To handle this, a custom backend service redirects BitTorrent client requests to the correct Hugging Face endpoint, differentiating between files stored in Git LFS and regular files. The project is still experimental—occasional HF CDN errors occur but are resolved with retries.

Planned enhancements include fully automating the process: generating .torrent files for new models, updating the website, and publishing all using GitHub Actions. However, GitHub's free runners only offer ~100 GB of disk space, posing a challenge for models exceeding that size. The developer is exploring workarounds like chunked torrents or external storage. This project tackles a growing need: reliable, decentralized distribution of large AI models (often tens to hundreds of gigabytes) without single points of failure or bandwidth bottlenecks.