An open-source registry called Model Registry now publishes AI model weights as torrents, a file-sharing method where machines trade pieces instead of pulling a whole file from a server. Each listing shows the size, licence, and live seed and peer counts. A daily job picks one model with a permissive licence, hashes it, and commits the torrent. More than 80 are listed, from GPT-2 to an 188 GB checkpoint. Each torrent also carries a web seed, a plain web address to pull pieces from, so downloads do not stall with no peer online. Mistral shipped Mixtral’s weights as a magnet link, a hash-based pointer for torrent clients, and it still sits in the official model card.
Model files have outgrown what one host serves comfortably. A 55 GB checkpoint a team pulls repeatedly is a fixed cost and a single point of failure, and a torrent spreads that work across everyone who already holds the file, a seed in torrent terms. The trade-off shows live on the site: many entries list zero seeds, so the web seed or the original host still carries most of the traffic. For now this is a spare mirror, not a swarm.
The other half of the idea stalled. Petals, a 2023 MIT-licensed project, split a model across volunteers so anyone could run Llama at home, BitTorrent-style, at up to 6 tokens per second on a 70B model. Its last code change was in August 2024. Research into decentralised serving continues, but in practice the file travelled and the compute did not.
Read More: The hedge against a central host is holding the weights yourself
Sources:
- Petals: run LLMs at home, BitTorrent-style (GitHub)
- Model Registry: decentralised distribution of open-source AI models
- modelregistry: the code behind the registry, including the daily torrent job (GitHub)
- PlanetServe: a decentralised overlay for large language model serving (arXiv)
- Mixtral 8x7B v0.1 model card, with the original torrent magnet link (Hugging Face)
Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service
Reuse
Citation
@misc{kabui2026,
author = {{Kabui, Charles}},
title = {BitTorrent for {AI:} {Model} {Weights} {Are} {Now} {Moving}
as {Torrents}},
date = {2026-09-14},
url = {https://toknow.ai/posts/bittorrent-for-ai-p2p-model-weights/},
langid = {en-GB}
}