PigeonAI Beta
One language model, run by the devices that join it, each holding a share. Lend yours and earn PigeonSeed, or run a model on your own device alone.
GitHub · Play against it in PigeonChess
A transformer split across peers
- No peer holds the whole model. Its layers are cut into stages, each held by the two peers closest to the stage's key, on their own GPU (WebGPU). An answer passes through the stages peer to peer, in order, over a PeerPigeon mesh.
- It grows with the mesh. Each layer routes every token to 2 of its experts, and the mesh keeps one expert per layer for each peer holding parts, so more peers means a bigger model: about 63 million weights with 10 peers, about 318 million with 64.
- Fast answers. A small draft model on the asker's own GPU guesses several tokens ahead, and one trip through the mesh checks them all, keeping the text exactly as the big model alone would write it.
- Trained together. Peers train it on one shared collection of text, cut into pieces spread over the mesh. Each peer updates only its own weights, so no weights or gradients cross the network: only activations and their gradients.
- It outlives its peers. Holders keep their stages in the browser, and a stage whose holder leaves moves to the next peer.
Where it runs
PigeonAI uses local relays only, never a cloud relay. A PigeonAI server can hold parts of the model too: PigeonStack runs one on a Mac's GPU.
PigeonSeed, the credit for lending a device, is coming soon.