Cross-tab WebGPU

One model, shared by every open tab of this page.

THIS TABCONNECTINGChecking coordination…—
LEADERElectingWaiting for Web Lock
MODELWaiting—

Prompt

Qwen3 1.7B · thinking enabled · runs in your browser · Ctrl/⌘+Enter to send

Response

No prompt yet.

How it works

tab A ─┐
tab B ─┼─ Web Lock ─ leader ─ WebGPU worker
tab C ─┘       ↑          ↕
           failover   BroadcastChannel

The leader keeps the only GPU model. Other tabs send prompts and receive streamed text through BroadcastChannel. Closing the leader releases the lock; another tab reloads the model from local OPFS cache.