Running in this tab · no server · no API key
Six verbs. Thirty-four models. Zero servers.
Edgewise runs language, vision, speech, judging, embedding, image and forecasting models on the device: in the browser through WebGPU or WebAssembly, and in Bun or Node. Right now, it can transcribe a meeting
The playground
Pick a verb. Pick a model. Press run.
Every demo below calls the published library in your browser. Models download once from Hugging Face and stay cached. What you type, say or show stays on this page.
Verbs compose
A voice assistant you can interrupt.
Silero VAD hears you start talking, Moonshine transcribes, LFM2.5 thinks and calls tools, Kokoro answers sentence by sentence. Talk over it and it stops mid-word. Four models, all in this tab.
The registry
34 models, one API.
Each model declares its verb, inputs, sizes and devices. The badge shows whether this device can run it. Click one to try it.
Ship AI that works offline, costs nothing per call and never sees your users' data.
import { generate } from 'edgewise';
for await (const delta of generate({ model: 'text:default', input: 'Hello!' })) {
out.textContent += delta;
}