Posted by Caren Chang – Developer Relations Engineer, Joanna (Qiong) Huang – Software Engineer, and Chengji Yan – Software Engineer
The latest version of Gemini Nano, our most powerful multi-modal on-device model, just launched on the Pixel 10 device series and is now accessible through the ML Kit GenAI APIs. Integrate capabilities such as summarization, proofreading, rewriting, and image description directly into your apps.
With GenAI APIs we’re focused on giving you access to the latest version of Gemini Nano while providing consistent quality across devices and model upgrades. Here’s a sneak peak behind the scenes of some of the things we’ve done to achieve this.
Adapting GenAI APIs for the latest Gemini Nano
We want to make it as easy as possible for you to build AI powered features, using the most powerful models. To ensure GenAI APIs provide consistent quality across different model versions, we make many behind the scenes improvements including rigorous evals and adapter training.
- Evaluation pipeline: For each supported language, we prepare an evaluation dataset. We then benchmark the evals through a combination of: LLM-based raters, statistical metrics and human raters.
- Adapter training: With results from the evaluation pipeline, we then determine if we need to train feature-specific LoRA adapters to be deployed on top of the Gemini Nano base model. By shipping GenAI APIs with LoRA adapters, we ensure each API meets our quality bar regardless of the version of Gemini Nano running on a device.
The latest Gemini Nano performance
One area we’re excited about is how this updated version of Gemini Nano pushes performance even higher, especially the prefix speed – that is how fast the model processes input.
For example, here are results when running text-to-text and image-to-text benchmarks on a Pixel 10 Pro.
| Prefix Speed – Gemini nano-v2 on Pixel 9 Pro | Prefix Speed – Gemini nano-v2* on Pixel 10 Pro | Prefix Speed – Gemini nano-v3 on Pixel 10 Pro | |
| Text-to-text | 510 tokens/second | 610 tokens/second | 940 tokens/second |
| Image-to-text | 510 tokens/second + 0.8 seconds for image encoding | 610 tokens/second + 0.7 seconds for image encoding | 940 tokens/second + 0.6 seconds for image encoding |
The future of Gemini Nano with GenAI APIs
As we continue to improve the Gemini Nano model, the team is committed to using the same process to ensure consistent and high quality results from GenAI APIs.
We hope this will significantly reduce the effort to integrate Gemini Nano in your Android apps while still allowing you to take full advantage of new versions and their improved capabilites.
Learn more about GenAI APIs
Start implementing GenAI APIs in your Android apps today with guidance from our official documentation and samples: GenAI API Catalog and ML Kit GenAI APIs quickstart samples.
