On-device AI means the model runs on your phone, tablet or laptop itself, so the text, photo or audio you give it is processed there instead of on a company’s server. Apple, Google and Samsung all document hybrid systems: some tasks stay on the device, and heavier ones go to a server. Whether a particular feature is “on-device” is a question about that feature, not about the phone.
The table below sets out what each platform’s own documentation says, as checked on November 3, 2026. After it come ways to see the route for yourself.
What runs locally on iPhone, Android and Galaxy phones
| Apple Intelligence (iOS 27) | Gemini Nano on Android (AICore) | Samsung Galaxy AI | |
|---|---|---|---|
| Local part | On-device models; Apple gives Priority Notifications as an example of a decision made by an on-device model | Gemini Nano, run by the AICore system service; ML Kit GenAI APIs process input, inference and output on the device | Most translation, Gallery edit suggestions, time and weather Ambient wallpapers, Google Messages writing suggestions |
| Server part | Private Cloud Compute (PCC) for requests that need larger models, such as Image Playground | Apps can add cloud models through “hybrid inference” when Nano’s limits are reached | Samsung lists auto formatting, auto summarizing, auto corrections and image editing as cloud-based |
| Can you see or force the route? | See: export the Apple Intelligence & PCC Report. Force: no single switch | No user switch; the route depends on how each app is built | Yes: Settings > Galaxy AI > Process data only on device |
| Hardware | iPhone 15 Pro, 15 Pro Max, iPhone 16 or later, iPhone Air, iPhone Duo | Device list per API; summarization and rewriting include Pixel 9, 10 and 11 and Galaxy S25 and S26 | Samsung’s list of supported Galaxy S, Z and Tab models; S22 and S23 series need a software update |
| Storage or other limits | Up to 8 GB, or up to 14 GB on iPhone 17 Pro models and later, iPhone Air and iPhone Duo; device and Siri language must match | AI inference only while the app is in the foreground | Samsung and Google account sign-in needed for the full set |
Sources: Apple’s Apple Intelligence requirements ↗, Apple Intelligence & Privacy ↗, Android Developers: Gemini Nano ↗, ML Kit GenAI APIs ↗ and Samsung: Use features with Galaxy AI ↗.
The table describes how each vendor characterizes its own system. It is not an independent audit, and lists change with software updates.
Apple: on-device first, Private Cloud Compute for the rest
Apple’s privacy page for Apple Intelligence ↗ says models run entirely on the device “in many cases”, and that requests needing more computing power can go to Private Cloud Compute, a server system built on Apple silicon. Its own example: when you use Image Playground, the device sends your description to PCC, where a server model creates the image. Apple says data processed in PCC is not stored or made accessible to Apple and is used only to fulfil the request.
A second clue sits on the requirements page ↗. Apple says features “that rely on server-side models” have daily usage limits, and names Siri AI, Intelligent Photo Editing Tools, Image Playground and AFM 3 Cloud models in Shortcuts. If a feature can hit a daily limit, at least part of it runs on a server.
The local part also takes room. The same page lists up to 8 GB of storage for most supported iPhones and up to 14 GB on iPhone 17 Pro models and later, iPhone Air and iPhone Duo. Apple Intelligence also requires the device language and the Siri language to be set to the same supported language.
Android: Gemini Nano runs inside a sealed system service
On Android, the on-device model is Gemini Nano, which runs in a system service called AICore. Google’s developer documentation ↗ says AICore has no direct internet access: model downloads go through a separate open-source component, Private Compute Services. Google also says AICore keeps no record of the input or output after a request.
Apps reach Nano through the ML Kit GenAI APIs ↗, which Google says run entirely on the device and keep working without a reliable connection. The summarization, proofreading, rewriting and image description APIs are listed for Pixel 9, 10 and 11 models, Galaxy S25 and S26 models and a range of other Android phones. An app may only run this inference while it is in the foreground.
Google’s documentation also tells developers to use hybrid inference when Nano’s token limits or capabilities are not enough. An Android app can therefore mix local and cloud processing in a way the phone itself does not reveal. Check the app’s own documentation or privacy policy.
Samsung: one switch that keeps Galaxy AI local
Samsung’s Galaxy AI support page ↗ is the most direct of the three. It says most translations and Gallery edit suggestions run on the device, and that most other features need a network connection and use the cloud. It names auto formatting, auto summarizing, auto corrections and image editing as cloud-based.
It also documents what it calls a master switch to disable all cloud-based processing for Galaxy AI: Settings > Galaxy AI > Process data only on device. While it is on, features that need the cloud are unavailable. Samsung says you can also allow only on-device processing feature by feature.
Worked example: summarizing a meeting note
Take one task, summarizing a long note, and follow it on each platform.
- Galaxy phone, Samsung Notes: auto summarizing is on Samsung’s cloud list. With “Process data only on device” turned on, that summary is not available, because cloud processing is off.
- Android app built on ML Kit: on a supported phone, the Summarization API runs locally through Gemini Nano. Google lists it as working without a reliable connection.
- iPhone: Apple doesn’t publish a per-feature route for summaries. Turn on the Apple Intelligence & PCC Report, run the summary and export the report. An entry means the request went to Private Cloud Compute; an empty report suggests it stayed on the phone. Apple notes the report may miss recent requests if Face ID or Touch ID is temporarily disabled.
The same request can stay on the device or reach a server, depending on the platform and the app.
How to check a feature yourself
- Read the vendor’s feature page, not the phone’s launch page. Requirements differ by feature, language and region.
- Use the platform’s report or switch. On iPhone, go to Settings > Privacy & Security > Apple Intelligence & PCC Report, choose a duration and export the file after using the feature. On a Galaxy phone, turn on Process data only on device and see which features stop working.
- Run an offline test. After any model downloads have finished, switch to airplane mode and try a short and a long request. A feature that waits, fails or shows an error needs the network for at least part of the job.
- Follow the result, not only the model. A summary made locally and saved to a cloud-synced note still leaves the phone. Our guide to files uploaded to AI assistants separates processing, retention and training.
Local processing does not make a result more accurate. Check a rewritten paragraph against what you meant and zoom in on an edited photo before you trust it. The AI testing guide gives a short, repeatable way to do that.
