You cannot ship on device inference with a real LLM still. Average people freak the hell out if it's even 10% slower than whatever google ships. At least thats the claim for how Firefox lost all their marketshare in the first place (Which was not the case. When everyone was claiming firefox was "slow", it simply was not, as long as you used an ad blocker. Normal people like my dad didn't switch to Chrome. It was installed through a sketchy mechanism and they never noticed)

Meanwhile, if you CAN run these models locally, using your own hardware for inference is supported. It's a setting option.