Train
Byrne70M, sized to this machine.
Same architecture as the Cycle2 checkpoint: 22 layers, hidden 576, grouped-query attention, three loops. The page measures this device and splits layers only when one machine cannot hold the weights.
No account
One model, your devices.
Train Byrne70M in the browser, or run a public model the same way. A computer holds the whole model. A phone takes the layers its memory can hold.
Infer
Public models only.
Each device downloads its own layers from Hugging Face. A gated or private repo is refused. There is no token.