Google has officially released its ML Kit GenAI API to Android developers, marking a significant step in advancing on-device generative AI capabilities.
Highlights
Quietly launched last week ahead of the Google I/O 2025 conference, this release provides access to Gemini Nano—Google’s compact, on-device AI model—enabling developers to build AI-powered Android apps without relying on cloud processing.
Previously, Gemini Nano was only available via limited experimental access through the AI Edge SDK introduced in October 2024.
That access did not permit developers to publish apps and was limited primarily to text-based tasks. With the introduction of the ML Kit GenAI API, developers can now integrate AI features into a broader range of Android devices, extending beyond the Pixel 9 series.
On-Device AI Features Now Available for Production
According to Google’s Android developer documentation, the API introduces a simplified, high-level interface designed to streamline the development of AI-driven app features.
These include message rewriting, proofreading, text summarization, and image description—all running directly on the device.
By executing tasks locally, the API enhances user privacy and reduces latency, thanks to its integration with AICore, a system service that leverages the device’s system-on-a-chip (SoC) for efficient model execution.
A notable addition is the image description capability, which was not previously accessible. This feature can be used to build applications that generate AI-based descriptions of images, improving usability for accessibility-focused tools and media organization platforms.
Multilingual Support and Limitations
The API currently supports a limited set of languages across its features:
- Text Summarization: English, Japanese, and Korean
- Proofreading and Rewriting: Supports seven languages, including German
- Image Description: Available initially in English only
Two Modes of Operation: Streaming and Non-Streaming
The ML Kit GenAI API supports both streaming and non-streaming modes:
- Streaming Mode: Delivers AI-generated output incrementally, useful for longer text or real-time feedback.
- Non-Streaming Mode: Returns the complete output after processing, suitable for shorter tasks or batch operations.
Built on AICore for Local AI Execution
The underlying framework for the API is AICore, an Android system service optimized for executing generative AI models like Gemini Nano directly on the device. This integration enables:
- Local Data Processing: Enhances privacy by keeping user data on-device
- Lower Latency: Improves speed and responsiveness
- Optimized Performance: Efficient use of device hardware via SoC integration
Expanded Device Compatibility
In contrast to the AI Edge SDK, which was restricted to the Pixel 9 series, the ML Kit GenAI API supports a wider array of Android devices.
This broadened compatibility allows more developers to experiment with and deploy on-device generative AI features, helping to accelerate adoption across the Android ecosystem.
Focus on Gemini Nano at Google I/O 2025
The timing of the API release appears to align with Google I/O 2025, where a session titled “Gemini Nano on Android: Building with On-Device Gen AI” is scheduled. During this session, Google is expected to:
- Showcase real-world use cases of the API
- Offer guidance for integrating Gemini Nano into Android applications
- Provide insight into the future roadmap of on-device generative AI