Can your phone run AI locally?

Yes — but only some of it, and your phone is quietly deciding which. The interesting question is not whether phones can run AI locally; it is which features stay on the device and which are sent to a server. That line matters, because anything that runs on-device is a task your data does not leave the phone to complete.

What "on-device" actually means here

Both big phone platforms now ship a small AI model that runs on the phone itself. These models are shrunk — quantized — to fit in a phone's memory, and they lean on the phone's NPU to run without flattening the battery.

On Android, it is Gemini Nano, Google's on-device model. It runs inside a system service called AICore and, in Google's words, "executes prompts locally, eliminating server calls". Apps reach it through Google's ML Kit for a fixed set of tasks — summarising, proofreading, rewriting a message in a different tone, describing an image, transcribing speech.

On iPhone, it is the ~3-billion-parameter on-device model behind Apple Intelligence. Apple says it "excels at" summarisation, text refinement and short dialog — and, tellingly, that it "is not designed to be a chatbot for general world knowledge". Anything bigger than those focused tasks goes to Apple's servers instead.

Notice the pattern: on-device models are small and good at narrow, well-defined jobs. The heavy, open-ended work — long conversations, image generation, questions that need broad world knowledge — still goes to the cloud.

Which feature runs where

A rough guide. Behaviour is vendor- and version-specific, and vendors move features across this line as their on-device models improve, so treat this as the current shape rather than a fixed rule.

How to tell which one you are using

You do not have to guess. Three checks, cheapest first:

1. Turn on airplane mode and try the feature. If it still works with no connection, it ran on-device. If it fails or asks you to go online, it needed the cloud. This is about the clearest signal you can get without developer tools.
2. Check whether your device even qualifies. On-device models need recent hardware — Apple Intelligence and Gemini Nano both list supported devices, and an older phone falls back to the cloud, or to nothing. If your phone is not on the list, "on-device" is not happening.
3. Watch for the vendor's own signal. Some features label cloud processing explicitly, or ask permission before sending data off the device. Read that prompt rather than clicking through it.

The payoff, and the limit

The reason to care is privacy: a summary generated on your phone was not uploaded, is not sitting in a server log, and is not feeding anyone's next training run. Google notes that AICore "doesn't store any record of the input data or the resulting outputs"; the on-device path is the one that keeps your data local by default.

The limit is capability. On-device models are small — a few billion parameters, with a modest token budget for context — so they are built for focused tasks, not the open-ended reasoning you would get from a large cloud model. If a feature feels sharper than your phone should manage, that is usually the tell that it went to the cloud. The practical move is to match the tool to the task: keep the private, bounded jobs on-device, and know that the impressive open-ended ones are borrowing a server.