Developers can now test Google’s Gemini Nano AI model on Pixel 9 devices

Gemini Nano is a more specialized, lightweight version of the broader Gemini family of AI models. While both belong to the same Gemini lineup, the Nano version is designed specifically for on-device AI tasks, that claims to make conversations more efficient and private, simply by processing data locally on the user's device.

Google has now opened Gemini Nano for experimentation on Android devices using its AI Edge SDK through AIcore. The on-device AI model claims to be optimized for tasks such as text generation and rephrasing. Following this announcement, developers will get a chance to test AI functionalities without relying heavily on cloud resources. However, there's a catch — it will initially only be available for Pixel 9 series devices.

Gemini Nano shines especially in text-based AI use cases like summarization, proofreading, and smart reply generation. The AI Edge SDK will let developers control specific parameters such as temperature, top-K sampling, and output length to fine-tune the model’s responses. For reference, top-K sampling is a method in AI models that limits the number of potential next words to the K most likely ones, trying to achieve a balance between coherence and randomness. This approach helps generate more relevant and less repetitive text while still allowing for variability in responses — all while minimizing the need for high-powered servers.

In terms of performance, Gemini Nano claims to have solid improvement over its predecessor. The model's accuracy in tasks like paraphrasing and math problems has jumped to 90% and 23%, respectively, as the company states. Google is currently offering experimental access to developers, letting them integrate all the aforementioned features into their apps.

To get started, developers can refer to the SDK’s documentation, which provides a step-by-step guide on using Gemini Nano in mobile apps.