App development in 2026: on-device AI, hybrid models and what to build first
Apple and Google have both opened their on-device AI models to developers. For businesses planning an app, that changes what is affordable, private and fast. What changed, and how we build with it.
Key takeaways
- At WWDC 2026 Apple expanded its Foundation Models framework, including free Private Cloud Compute access for smaller developers and image input (MacRumors, June 2026).
- Android developers can use Gemini Nano on-device through ML Kit GenAI APIs for summarising, proofreading, rewriting and describing images (Android Developers Blog).
- The practical pattern for 2026 is hybrid: on-device AI for fast, private tasks, and cloud models for heavier reasoning, with a fallback for older phones.
Apple: Foundation Models grows up
At its June 2026 developer conference, Apple outlined a significant expansion of its AI tools for app makers. According to MacRumors, developers with fewer than two million first-time App Store downloads get free access to Apple's Foundation Models running on Private Cloud Compute, removing server costs for smaller apps. The framework now accepts image input, can call third-party server models through the same Swift API, and Apple said it would open-source the framework.
Apple also introduced Core AI, a framework for running custom models on Apple silicon, with tools to convert PyTorch models. Callstack notes that Apple is adding a more capable on-device model for higher-end devices, which means features can now behave differently depending on the phone in someone's hand.
Android: Gemini Nano for app developers
On Android, Google offers Gemini Nano through ML Kit GenAI APIs for summarisation, proofreading, rewriting and image description, running on the device. Google reported that the latest version on the Pixel 10 Pro processes text almost twice as fast as the previous generation on the Pixel 9 Pro.
The trends that matter for businesses
- Hybrid AI by default. Quick, private tasks run on the phone; heavier reasoning goes to a cloud model. Users get speed and privacy, and you control cost.
- Design for device tiers. On-device AI is not available on every phone. Every AI feature needs a graceful fallback.
- Cross-platform still wins for most apps. One codebase for Android, iOS and the web keeps cost and maintenance down. Native code is reserved for features that need it.
- System integration. Apps that expose their actions and content to the operating system become easier to reach through assistants and search.
- Offline-first. Apps that keep working without a connection feel faster and more trustworthy.
How we built this into our own apps
These are not abstract trends for us. GameHarbor, our wellbeing app, runs as one JavaScript codebase on Android, iOS and the web. Its companion, Ollie, has an on-device version that works fully offline, with optional cloud AI replies when a connection is available, and speech is handled on the device. MindHarbor uses Three.js for a real-time 3D world packaged for mobile with the same cross-platform approach.
The lesson from both projects is to decide early which features must work offline and privately, and which can depend on the cloud. That decision shapes cost, speed and user trust more than any choice of framework.
Where should each AI feature run?
| Feature | On-device | Cloud |
|---|---|---|
| Summarising or rewriting a user's own text | Good fit: fast and private | Fallback for unsupported phones |
| Voice input for short commands | Good fit, and works offline | Only for long or complex speech |
| Conversational companion or assistant | Short replies and offline mode | Longer, more capable conversations |
| Search across large product or content catalogues | Rarely | Usually, with the catalogue on the server |
| Image understanding | Simple descriptions on supported devices | Detailed analysis |
The pattern is consistent: keep personal, latency-sensitive tasks on the phone, send heavy or catalogue-wide tasks to the cloud, and always define what happens when neither is available.
What drives cost and timeline
- Number of platforms. Cross-platform reduces this, but each store still needs its own listing, review and testing.
- Accounts and data. Sign-in, sync and admin tools add more work than most people expect.
- AI features. Each one needs prompts or models, evaluation, safety rules and a fallback path.
- Content. Text, images, sound and in-app copy are often the slowest part of a launch.
- Maintenance. New OS versions and model updates arrive every year; budget for them from the start.
Planning an app in 2026
- Start with one core job the app must do well, and prototype it with real users.
- List every AI feature and decide: on-device, cloud or both, and what happens on older phones.
- Choose cross-platform unless a specific feature demands native code.
- Plan privacy from the start: what stays on the device, what is stored, and how users control it.
- Budget for after launch: updates for new OS versions, model changes and store reviews.
If you are planning an app, our app development team can help you scope it.
Questions and answers
What is on-device AI?
On-device AI runs a model directly on the phone rather than on a remote server. It is faster for short tasks, works offline and keeps data on the device, but device models are smaller than cloud models.
What did Apple announce for developers at WWDC 2026?
An expanded Foundation Models framework with image input, the ability to call third-party server models through the same API, free Private Cloud Compute access for developers under two million first-time downloads, and a new Core AI framework for running custom models on Apple silicon.
Can Android apps use Gemini Nano?
Yes. Google offers Gemini Nano through ML Kit GenAI APIs for on-device summarisation, proofreading, rewriting and image description on supported devices.
Should I build a native or cross-platform app in 2026?
For most business apps, cross-platform: one codebase for Android, iOS and the web. Native development is worth it when a feature depends on platform-specific capabilities.
