Phones now advertise on-device intelligence, yet many requests still travel to a data center. The split is decided by resource limits rather than by policy preference.
Memory is the first constraint
A model must fit in the memory available alongside everything else the phone is running. Phone memory is a fraction of what a server allocates to a single model.
Compression techniques shrink models substantially, at some cost to accuracy, and that trade determines what can run locally at all.
The practical outcome is that small, narrowly scoped models run on the device while large general ones do not.
Power and heat cap sustained work
Running a model continuously draws power and produces heat, and a phone must stay comfortable to hold and last a day.
Short tasks fit inside that budget easily. Long generations do not, and the system will throttle rather than let temperature rise.
Server execution moves that cost off the device entirely, which is why longer or repeated requests are the ones most likely to be routed away.
The routing decision is made per request
Systems evaluate what the request needs: whether it requires current information, how much context it involves, and whether a local model can serve it acceptably.
That evaluation happens quickly and invisibly, which is why two similar-looking requests can behave differently in latency and in offline availability.
Some platforms surface the distinction with an indicator or a setting, but the default is to optimize for a usable answer rather than for locality.
Privacy claims attach to the handling, not the location
Server processing does not automatically mean data retention. Several architectures process requests in isolated environments and discard them afterward.
Verifying that requires trusting published designs and audits, whereas on-device processing can be verified more simply by observing that no network traffic occurs.
The meaningful question for a user is therefore what is retained and for how long, which is stated in policy rather than implied by architecture.
Offline behavior reveals the split
Turning off connectivity is the clearest test. Features that keep working are local; features that fail or degrade were being served remotely.
Transcription, keyboard prediction and photo subject detection commonly survive. Open-ended writing and question answering usually do not.
That boundary shifts as memory grows and models shrink, so the same feature can migrate on-device across a software update without any announcement.