The Model Bay

What this is

The Model Bay keeps open model weights shared even after the original host takes them down. The client resolves a model to verified bytes via a signed index, fetches over a BitTorrent v2 swarm (with Hugging Face as the source until takedown), checks every byte against the published hashes, lays files out so from_pretrained works, and then seeds.

Using the client

curl -fsSL https://themodelbay.org/install.sh | sh

macOS today; Windows, Linux, x86 and full source coming soon.

modelbay index list                              # browse the curated models
modelbay pull "<model>" --dest ./models           # verify, lay out, and seed
modelbay pull "<model>" --dest ./models --no-seed  # pull without seeding
modelbay pull "<model>" --dest ./models --torrent-only  # swarm only, no Hugging Face
modelbay seed "<model>" --dest ./models           # seed a model you already pulled
modelbay pull "<model>" --dest ./models --index ./index.json --trust ./trust.json  # use another list

Pulled models land in the Hugging Face cache layout, ready for from_pretrained / vLLM. --index points the client at a different signed index (a fork you mirror, or a third-party publisher's list) and --trust gives the public keys to verify that index against, instead of the curated index built into the client.

Using the site

Browse the curated models, copy the install line, or copy a model's pull command. Open a model for its infohash, magnet link, .torrent, and file manifest. The site verifies the index signatures in your browser; it does not download or host model bytes. Your client does that, and re-checks every byte against the signed hashes on pull.

Trust

The index is signed; the client trusts these pinned public keys. A host that swaps the index cannot forge entries without the private keys.

The index is a signed static repo at https://github.com/SirFrancisWeights/the-model-bay-index. Clone it to fork or mirror; the trust root is the public key, not the host.