I think apple published that study right before WWDC to have an excuse to not give bigger than 3B foundation models locally and force you to go via their cloud -for reasoning- harder tasks.

beta api's so its moving waters but that's my thoughts after playing with it, the paper makes much more sense in that context