Running a model on your phone

A recent phone runs real language models, offline, today — the same open-weight files a computer runs, in their smallest quantized sizes. It takes an app and a download, and knowing what to expect.

The constraint is stricter than the spec sheet. A phone never gives one app all its memory: the system enforces a per-app ceiling well below the total, and crossing it doesn't slow the app, it kills it. This is why our phone pages reason from the per-app ceiling, labelled as the estimate it is, rather than the number on the box. The practical result: today's phones run models in the 1–4B class comfortably, and the fit verdict on each phone's page tells you which.

What that class of model is honestly good for. Drafting and rewriting, summarising what you paste, notes and lists, questions with self-contained answers, translation that beats being offline in a foreign country. What it is not: a pocket frontier model — reasoning depth and factual reach drop steeply at this size, and the first hallucinated fact should surprise you less than it will. Offline and private is the trade, not free capability.

Getting one running. App stores on both platforms carry free apps built on the standard local runtimes — search for "local LLM" or "offline AI", pick one that lets you choose the model rather than hiding it, and download a small model from within the app. First generation after opening is slow while the model loads; after that it types at a readable pace on a recent device. Expect warmth and battery drain during long sessions — reading the whole model per token is the same physics as everywhere else, on a chip built for efficiency.

The quiet best use. A phone model pairs naturally with local transcription: capture thoughts or meetings on the device, summarise on the device, and nothing ever leaves it. For anything heavier, your phone is also the perfect client for a model served from a Mac mini at home — the phone does the talking, the desk machine does the thinking.

Where next: Phones in the hardware index · What fits and why · Capturing meetings