Android 17 Beta unveils native generative AI APIs and on-device sandboxing

AI · 5 min read

Android 17’s developer preview includes a set of high-level generative AI APIs that let apps request text, image, and audio generation with standardized prompts and response formats. The APIs abstract tricky tasks like prompt shaping and multilingual output, accelerating integration for apps across categories.

To address performance and safety, Google introduced an on-device model sandbox that isolates third-party models from the OS and other apps while providing hardware-accelerated inference. The sandbox enforces resource limits and a permission model for sensor and network access so apps can run powerful models locally without risking data leakage.

Google also expanded the Neural Networks API to better leverage mobile GPUs and dedicated NPUs, and added tools in Android Studio to profile model latency and energy use. The company emphasized that the APIs are optional and that developers can still call cloud-based models when heavier compute is required.