When a card is generated for the first time, the browser downloads approximately 700 MB of model files and caches them locally on the device; all subsequent inference is performed locally on the device without going through a server.