All inference on-device you mean? I presume that the vast majority of their models are trained in the cloud. Apart from some fine-tuning for personalization maybe.

While they did say machine learning, all the examples are of inference, and not training. So you don't have to ask if they meant inference when it is very clear what they mean.

Outside of marketing speak, it is not really machine learning if the machine is not learning.

I wouldn't expect "this phone can run for 10 hours and train a model", even though it clearly is also possible.

The recognition of faces and settings of images for example is fully done on the phone. But that is done overnight when the phone is plugged in.