Gemini Nano is now available on Android via experimental access

Google introduced Gemini, a family of multimodal models designed for flexibility and capable of running on various devices. Gemini Nano is the most efficient model, suitable for on-device tasks. It has been made available to all Android developers through the AI Edge SDK using AICore. Developers can experiment with text-to-text prompts on Pixel 9 series devices, with support for more devices and modalities coming in the future. Gemini Nano excels in academic benchmarks and real-world applications, offering capabilities that rival much larger models. It is already used in Google apps such as Talkback and Pixel Recorder. Integrating AI models into mobile apps can be challenging due to computational resources and storage space requirements. AICore, a new system service in Android, allows developers to benefit from AI running directly on the device without needing to distribute runtimes, models, and other components themselves. The AI Edge SDK enables developers to customize prompts and inference parameters to their specific needs, providing greater control over each inference. Developers can experiment with Gemini Nano by adding a specific dependency to their apps and customizing inference parameters like temperature, top K, candidate count, and max output tokens. A guide on integration and prompting strategies is available for developers to start experimenting with Gemini Nano.