Posted by Caren Chang – Developer Relations Engineer, Joanna (Qiong) Huang – Software program Engineer, and Chengji Yan – Software program Engineer
The newest model of Gemini Nano, our strongest multi-modal on-device mannequin, simply launched on the Pixel 10 machine collection and is now accessible by way of the ML Equipment GenAI APIs. Combine capabilities equivalent to summarization, proofreading, rewriting, and picture description immediately into your apps.
With GenAI APIs we’re centered on providing you with entry to the most recent model of Gemini Nano whereas offering constant high quality throughout gadgets and mannequin upgrades. Right here’s a sneak peak behind the scenes of among the issues we’ve finished to attain this.
Adapting GenAI APIs for the most recent Gemini Nano
We wish to make it as simple as doable so that you can construct AI powered options, utilizing probably the most highly effective fashions. To make sure GenAI APIs present constant high quality throughout completely different mannequin variations, we make many behind the scenes enhancements together with rigorous evals and adapter coaching.
- Analysis pipeline: For every supported language, we put together an analysis dataset. We then benchmark the evals by way of a mixture of: LLM-based raters, statistical metrics and human raters.
- Adapter coaching: With outcomes from the analysis pipeline, we then decide if we have to prepare feature-specific LoRA adapters to be deployed on prime of the Gemini Nano base mannequin. By transport GenAI APIs with LoRA adapters, we guarantee every API meets our high quality bar whatever the model of Gemini Nano working on a tool.
The newest Gemini Nano efficiency
One space we’re enthusiastic about is how this up to date model of Gemini Nano pushes efficiency even greater, particularly the prefix velocity – that’s how briskly the mannequin processes enter.
For instance, listed below are outcomes when working text-to-text and image-to-text benchmarks on a Pixel 10 Professional.
| Prefix Velocity – Gemini nano-v2 on Pixel 9 Professional | Prefix Velocity – Gemini nano-v2* on Pixel 10 Professional | Prefix Velocity – Gemini nano-v3 on Pixel 10 Professional | |
| Textual content-to-text | 510 tokens/second | 610 tokens/second | 940 tokens/second |
| Picture-to-text | 510 tokens/second + 0.8 seconds for picture encoding | 610 tokens/second + 0.7 seconds for picture encoding | 940 tokens/second + 0.6 seconds for picture encoding |
The way forward for Gemini Nano with GenAI APIs
As we proceed to enhance the Gemini Nano mannequin, the staff is dedicated to utilizing the identical course of to make sure constant and prime quality outcomes from GenAI APIs.
We hope it will considerably scale back the hassle to combine Gemini Nano in your Android apps whereas nonetheless permitting you to take full benefit of recent variations and their improved capabilites.
Be taught extra about GenAI APIs
Begin implementing GenAI APIs in your Android apps at the moment with steering from our official documentation and samples: GenAI API Catalog and ML Equipment GenAI APIs quickstart samples.
