Until recently, adding AI to an app meant sending user data to a cloud service and paying for every request. Now, modern phones can run capable AI models locally. Apple opened its on-device foundation model to developers with iOS 26, and Google offers Gemini Nano on many Android devices. That opens new kinds of features: private, offline and free to run at scale.
Here's what's possible and how to decide between on-device and cloud AI.
What's available
| Platform | Options |
|---|---|
| iOS | Apple's Foundation Models framework (on-device language model, iOS 26+ on Apple Intelligence devices); Core ML for custom models; Vision, Speech and Natural Language frameworks |
| Android | Gemini Nano through Google's on-device AI APIs on supported devices; ML Kit for vision, text and translation; LiteRT (formerly TensorFlow Lite) and other runtimes for custom models |
| Cross-platform | Custom models via ONNX Runtime or similar; Flutter and React Native plugins for platform AI features |
What on-device AI does well
- Summarising notes, messages and articles.
- Classifying and tagging: sorting expenses, labelling photos, routing messages.
- Extracting information: dates, amounts and names from text or receipts.
- Rewriting and proofreading short text.
- Smart replies and suggestions.
- Text recognition (OCR) from the camera.
- Translation for common languages.
- Speech to text for dictation and voice commands.
- Image tasks: background removal, object detection, simple enhancements.
Benefits
| Benefit | Why it matters |
|---|---|
| Privacy | Personal data stays on the phone |
| Offline | Works on planes, underground and in poor coverage |
| Speed | No network round trip for small tasks |
| Cost | No per-request cloud fees, however many users you have |
| Compliance | Less personal data processed on servers. See using AI without breaking GDPR |
Limits
- Smaller models: less general knowledge and weaker complex reasoning than large cloud models.
- Device support: only recent, higher-end phones run on-device generative models, so many users won't have them.
- Battery and heat: heavy use drains batteries.
- App size: bundling custom models increases download size.
- Updates: improving the model may depend on OS updates or app releases.
On-device vs cloud: how to choose
| Feature need | Better fit |
|---|---|
| Sensitive personal data (health, finance, private notes) | On-device |
| Must work offline | On-device |
| Short, focused tasks at high volume | On-device |
| Broad knowledge or up-to-date information | Cloud |
| Complex reasoning, long documents, coding | Cloud |
| Must work the same on every phone | Cloud, or on-device with cloud fallback |
Many good apps use a hybrid: try on-device first, fall back to the cloud when the device can't handle it or the task is too complex, with clear user consent for anything sent to servers.
Feature ideas by industry
| Industry | On-device AI feature |
|---|---|
| Finance and bookkeeping | Read receipts and suggest categories offline |
| Health and fitness | Summarise private journal entries without uploading them |
| Field services | Dictate job notes offline and extract parts used |
| Retail | Scan a label and translate it; visual product search |
| Education | Summarise lesson notes; generate practice questions |
| Productivity | Smart replies, meeting note summaries, task extraction |
Design tips
- Label AI features and let users edit results. See designing AI features users trust.
- Detect device capability and hide or replace features gracefully on unsupported phones.
- Keep tasks small and focused to suit on-device models.
- Measure quality on real devices, not just simulators.
- Explain privacy clearly: "This runs on your phone; your notes aren't uploaded" is a strong selling point.
A hybrid architecture that works
A common pattern for production apps:
- Check capability at startup: does this device support the on-device model?
- Route by task: short, private tasks (summarise a note, extract an amount from a receipt) go on-device when available.
- Fall back to the cloud for unsupported devices or complex tasks, with clear consent and minimal data sent.
- Cache results so the same work isn't repeated.
- Measure quality for both paths, and keep prompts and outputs consistent so users get similar results either way.
Example: receipt capture in an expenses app
| Step | Where it runs |
|---|---|
| Photograph receipt and read the text (OCR) | On-device |
| Extract supplier, date, total and VAT | On-device model where supported; cloud fallback |
| Suggest an expense category | On-device, using the user's past categories |
| Sync the final expense to accounts | Server (only the structured result, not the photo, if that's all you need) |
Users get instant results offline, their receipt photos needn't leave the phone, and your cloud AI bill only covers devices that can't do the work themselves.
Cost implications
On-device AI moves cost from ongoing cloud fees to upfront development and testing. For apps with many users and frequent AI use, that can save a lot. For features that need large models, cloud APIs remain necessary. Plan both in your budget. See how much a mobile app costs in 2026 and app maintenance costs.
The short version
Phones can now run AI models locally through Apple's Foundation Models framework, Google's Gemini Nano and custom models. On-device AI is private, offline, fast and free per request, and great for summarising, classifying, extracting and rewriting. It's limited to newer devices and smaller tasks, so many apps combine on-device AI with a cloud fallback.
Want AI features in your app? Our mobile app development team and AI automation team build on-device and hybrid AI features.




