They are playing the long game, which is what has always been: wait until the silicon enables that for most users. The Apple Silicon track record suggests that... wait a couple of years and we'll get M3-Ultra-class capabilities in all of Apple devices. Some day the lowest bar will be above running state of the art LLMs on device.

Siri hasn't run on device for most of its existence. It's only in the last few years that Apple suddenly decided it was a priority.

All they have to show is incremental improvements over Siri. For that, Quantized models are more than enough in my opinion.