By "dust settled" I don't mean that the technology "exists" -- but rather that feature development has slowed down and most products have stabilized as feature complete and mature.
The on-device ML models that are being used by Google and Apple are both quite new and in active development.
Many of Apple's most successful products have shipped years or even a decade after their competitors. They have tried using first-mover advantage in the past but typically fail when using that strategy.
They're in active development, but they already worked well at launch in English in 2019, serving enough customers to be very useful. I was using it myself.