A 2B-parameter speech-to-speech omni model runs entirely on WebGPU in this
tab: mic in, synthesized voice out, no server. Every turn card shows the measured latency
chain: end-of-speech (how long the VAD took to decide you finished),
first word (model's first text token) and first voice (first
synthesized audio). That chain is where voice agents live or die.
1. Click Load model (one-time ~2.4 GB, then browser-cached). 2. Just talk — hands-free is on: it answers out loud. 3. Interrupt it by talking over it (barge-in). Headphones recommended (echo otherwise). The mic button still works as push-to-talk.