Running a language model on a phone stopped being a demo. The question is no longer whether it works, but which jobs it does well enough to be worth the battery, the storage and the heat.
What phones are good at
- Short summaries of a note, an email or a transcript.
- Rewriting a sentence in a different tone.
- Extraction: dates, amounts, names from a block of text.
- Offline dictation and translation of short passages.
All of these are narrow, local and latency-sensitive. That is exactly the profile where a small model on an NPU beats a round trip to a data centre.
What they are not good at
Long-document reasoning, multi-step planning, and anything that needs current information. A phone model has a small context window and no browsing. Ask it to compare five contracts and it will lose the thread, politely.
The practical constraints
- Storage: a quantised model is measured in gigabytes, and each additional model is another download.
- Memory: the model must fit alongside the app and the system, so background apps get killed.
- Heat: sustained generation warms the chassis quickly; long jobs throttle.
- Battery: an hour of generation is a visible line in the battery graph.
How to try it sensibly
Start with one small model and one app that clearly states it runs locally. Test it in airplane mode so you know nothing is leaving the device. Measure how long a typical summary takes and how warm the phone gets. If it is uncomfortable after ten minutes, the use case is not a phone use case.
Where this is heading
The direction is clear: more capability per gigabyte, better NPU scheduling, and system-level features that call a local model without an app. What will not change is the ceiling: a phone is a thermally limited device, and long reasoning will stay a server job for a while.
Bottom line: use a phone model for small, private, instant tasks. Reach for the cloud when the task needs breadth, and do not expect the gap to close this year.
