Posted by Dave Burke, VP of Engineering
Rather than being trained for a narrow use case, Foundation models learn from a variety of data sources to produce AI systems that can adapt to different tasks. We’ve just announced our most capable model yet, the Gemini. Gemini is designed for flexibility, so it can work in everything from data centers to mobile devices. Optimized for three different sizes: Ultra, Pro and Nano.
Gemini Nano, optimized for mobile
Gemini Nano, our most efficient model built for on-device operations, runs directly on mobile silicon, unlocking support for a variety of important use cases. On-device processing enables features where the data does not leave the device, such as suggesting replies to messages in an end-to-end encrypted messaging application. It also enables consistent experiences on latency, so features are always available when there’s no network.
Gemini Nano is a successor to the larger Gemini models and is specially optimized to run on mobile silicon accelerators. Gemini Nano enables powerful capabilities such as high-quality text summarization, contextually intelligent responses, and advanced proofreading and grammar correction. For example, the Gemini Nano’s improved language understanding allows the Pixel 8 Pro to summarize content in a recording app, even when the phone’s network connection is offline.
Gemini Nano has started enabling Smart Response in Gboard on the Pixel 8 Pro to be enabled as a developer preview in Settings. Available now to test on WhatsApp and coming to more apps next year, the on-device AI model saves you time by suggesting high-quality replies with speech understanding.1.
Android AICore, a new system service for the base models on the device
Android AICore is a new system service available in Android 14 that provides easy access to Gemini Nano. AICore handles model management, runtime, security features and more, making it easy to incorporate AI into your applications.
AICore is private by design, following the example of Android’s private compute core and providing transparency and auditability through open source APIs. As part of our efforts to build and deploy AI responsibly, we’ve built special security features to make it safer and more accessible to everyone.
AICore Low Level Alignment (LoRA) enables fine-tuning with Gemini Nano. This powerful concept allows application developers to create small LoRA adapters based on their own training data. The LoRA adapter is powered by AICore, resulting in a powerful large language model tailored to the application’s own use cases.
AICore uses the latest ML hardware such as the latest Google Tensor TPU and NPUs by leading Qualcomm Technologies, Samsung S.LSI and MediaTek silicon. AICore and Gemini Nano are rolling out to the Pixel 8 Pro, with additional devices and silicon partners to be announced in the coming months.
Build in Gemini
We’re excited to bring together state-of-the-art AI research with easy-to-use tools and APIs for Android developers to build with Gemini. If you are interested in building applications using Gemini Nano and AICore, please sign up for our Early Access Program.