Pooled
A browser-based system that pools the GPUs of several devices over WebRTC to run one large open language model together, splitting its layers across laptops, desktops and phones.
Pooled runs one open language model across several devices in a "room": a laptop, a friend's desktop and a phone can each hold some of the model's layers and run them on their own GPU using custom WGSL kernels, combining into more capacity than any single device has alone. Each token's hidden state passes between devices over direct WebRTC connections, and the host assembles the final result.
It needs no install, no account and no server-side inference; everything runs in browser tabs on the same Wi-Fi network. It supports ordinary chat as well as a Code mode for building and fixing small apps, and includes speculative decoding, batched prefill and golden-tested kernels so distributed decoding matches plain single-device decoding.