Your phone will not give one app all its memory. The system holds each app to a ceiling well below the number printed on the box, and an app that reaches past it gets closed rather than slowed down, which is why a model that looked like a comfortable fit dies three messages into a long conversation. Open any handset in our phone pages and that is the figure at the top: what one app may use, beside what the device holds, marked as the estimate it is. The iPhone 17 Pro's page is the clearest example of the gap.
Work from that ceiling and the answer comes out at models of roughly one to four billion parameters, stored in their smallest compressed sizes. Those are real models with openly published weights, the same files a laptop runs, and each phone page lists which of them fit it and how fast they are likely to be.
What that class does well is narrower than the marketing around phone AI suggests, and worth knowing before you download anything. It drafts and rewrites. It summarises text you paste into it. It keeps notes and lists in order. It answers questions that contain their own answer somewhere in the question. And it translates, which is the job where a phone model beats everything else you have, because the moment you need it most is usually the moment you have no signal. Reasoning depth and factual reach fall away steeply at this size, so the first confidently wrong answer should arrive sooner than you expect.
Getting one running takes about ten minutes. PocketPal AI is the one to install first: it is free, it runs on both iPhone and Android, and it lets you choose the model and see how big it is instead of hiding both behind a chat window. MLC Chat, from the project behind much of the on-device toolchain, does the same job with a rougher edge and is worth trying if PocketPal will not run something you want. Either way you pick a model inside the app and it downloads. The first reply after opening is slow while the model loads, and after that it types at a readable pace on anything recent.
The best use of a phone model is the one nobody advertises, which is pairing it with local transcription. Record the meeting on the device, transcribe it on the device, summarise it on the device, and the recording of other people talking never goes anywhere. And when the job outgrows the phone, the phone makes an excellent client for a model served from a machine at home: the handset does the talking and the desk does the thinking.
Where next: Phones in the hardware index · What fits and why · Capturing meetings