On-device AI for Android Developers
On-device AI means running machine learning models directly on the phone instead of calling a server. Gemini Nano offers on-device generative features through Android system services, and ML Kit provides ready-made APIs for text, barcode, face and language tasks. Benefits include lower latency, offline use and keeping data on the device.
Stage: Release and Beyond · Part of the Android Developer roadmap · Last reviewed 19 November 2026
What to learn in On-device AI
- Explain when on-device inference is better than a cloud API
- Scan a barcode or recognise text using an ML Kit API
- Describe what Gemini Nano is and where its availability is limited
- Handle the case where a model is unavailable on a device
- Consider privacy and battery cost when adding ML features
On-device AI interview questions
Why run a model on the device instead of calling a cloud API?
On-device inference avoids network round trips, works offline, keeps personal data on the phone and has no per-request server cost. The tradeoffs are smaller models, limits from device memory and battery, and uneven hardware support, so you usually need a fallback path.
What is ML Kit?
ML Kit is Google's SDK of ready-to-use machine learning APIs for mobile, such as text recognition, barcode scanning, face detection and language identification. Most run on the device. You call them as normal libraries without training or hosting your own model.
Where On-device AI fits in the Android Developer roadmap
On-device AI belongs to the Release and Beyond stage (intermediate level), which takes about 6 hours. Ship and grow.
Other units in this stage
Back to the full Android Developer roadmap