Google has enhanced its on-device AI capabilities for Android apps through the expansion of Gemini Nano and ML Kit’s GenAI APIs, enabling more practical, data-secure local AI functions for developers amid broader industry adoption.
Android’s on-device artificial intelligence story has moved beyond novelty and into something developers can actually use. Google has expanded Gemini Nano access through ML Kit’s GenAI APIs, giving Android apps a way to handle tasks such as summarisation, proofreading, rewriting, image description and speech recognition locally on the device. That matters because it keeps data off the cloud, cuts latency and reduces dependence on a live internet connection, while AICore provides the system layer that helps run and update the model.
For app makers, the appeal is practical rather than theoretical. Android’s official guidance says Gemini Nano can deliver richer generative AI experiences without sending data to a server, and the ML Kit APIs sit on top of that foundation with higher-level entry points for common features. In effect, developers can choose a simpler route when their use case fits one of Google’s supported tasks, instead of building the whole inference pipeline themselves.
That is why rewriting is such a useful example. The ML Kit rewriting API is still in beta, requires Android 8.0 or later and supports a small set of output styles, including friendly, professional, shorten and emojify. The API also expects developers to check feature status first, because the model or feature-specific assets may need to be downloaded before inference is available.
Google’s documentation also highlights the limitations that come with on-device AI. Input is capped at 256 tokens, the supported languages are limited and the feature is unavailable on devices with an unlocked bootloader. That makes the UX work just as important as the AI itself: apps need loading states, fallbacks and clear error handling if they want the experience to feel dependable rather than experimental.
Where the built-in APIs are not enough, LiteRT remains the custom route. Android developers can convert TensorFlow models into a .tflite format for local use, but Google recommends checking compatibility carefully because not every TensorFlow operation is supported. SavedModel is the preferred export path, and some models may need refactoring or extra optimisation work before they run cleanly on device.
The broader trend is clear: on-device AI is becoming a mainstream part of Android development. Industry reporting and Google’s own materials point to Gemini Nano reaching a far wider device base, with specialised local models increasingly used for vision, text and audio tasks. For developers, that means the decision is less about whether on-device AI is possible and more about whether to start with ML Kit’s higher-level APIs or move straight to a fully custom model pipeline.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
