Where does the AI actually run?

On your device. Model weights download once into your browser's cache and inference happens locally in a Web Worker, on WebGPU where the browser supports it. The page's content-security policy contains no inference endpoint, so a call to a hosted model would be blocked before it left.

Is on-device AI as good as a cloud model?

Not on the hard cases. A model that has to fit in a browser cache is a fraction of the size of one running in a data centre, and the difference shows on fine hair, glass and motion blur. On ordinary photographs the gap is small enough not to matter.

Why is the first run slow?

Because the weights have to reach your machine before anything can happen — tens of megabytes, once. After that they are cached and the tool works offline, which a hosted model never does.

Does the summariser write new sentences?

Only when your browser provides an on-device language model. Otherwise it is extractive: it selects the most central sentences already in your text and returns them verbatim. That is less fluent than a written summary, and it is what it claims to be.

On-Device AI vs Cloud AI Tools — An Honest Look

Background removal, upscaling and proofreading that run on your machine, and a straight account of what a hosted model still does better.