WebGPU Cluster turns a browser tab into an AI inference host. Pick a model, keep the tab open, and a Node broker can send it detection or image-description jobs through a simple HTTP API.
The current build supports RF-DETR Medium for object detection and SmolVLM-500M for image description, both loaded from Hugging Face and run in the host browser through WebGPU.
It is less about a polished consumer app and more about a useful question: how much shared compute can the web expose when WebGPU becomes the worker, not just the renderer?
