It's because of the local model use case. Running AI locally on the phone is a feature (they brag about the processor), but it's constrained by memory and bandwidth.