EXPERIMENT · DISTRIBUTED INFERENCE
What if the computer was the room? We are testing whether ordinary phones, laptops and tablets can pool their spare computing capacity and run a single AI model together. Each device holds part of the model; the room becomes the computer.
FIELD TEST · COLLECTIVE COMPUTE
One model. Several devices.
There is no model API doing the inference in the middle. Each browser downloads only its assigned model layers, computes them locally with WebGPU, then passes intermediate activations to the next participating device over WebRTC.
Nothing entered here should be private. Participating devices are part of the computation, and model activations should be treated as potentially recoverable by participants.
Bring 2+ modern devices
Runtime WebGPU + WebRTC
Quick model Qwen3 0.6B Q4 · ~429 MB total
Weights Downloaded directly from Hugging Face
Built on SwarmLLM · MIT