Following the exciting launch of Gemma 3 and Gemma 3 QAT, we are further driving our vision of accessible AI as we are a family of cutting-edge open models that can run on a single cloud or desktop accelerator. Gemma 3 provided powerful features to developers. We are now extending that vision to highly capable, real-time AI that works directly on the devices we use every day (cell phones, tablets, laptops).
We have designed new cutting-edge architectures to enhance next-generation on-vis AI, including by advancing the capabilities of Gemini Nano, and to support a wide range of applications. This next-generation foundation was created in close collaboration with mobile hardware leaders such as Qualcomm Technologies, MediaTek, and Samsung’s Systems LSI business, and is optimized for Lightning-Fast’s multimodal AI, allowing true personal and private experiences directly on the device.
The Gemma 3N is the first open model built on this groundbreaking, shared architecture, allowing developers to start experimenting with this technology in today’s preview. The same advanced architecture drives the next generation of Gemini Nanos. This brings these features to the wide range of features in Google Apps and our device ecosystem, and will be available later this year. With Gemma 3N you can start building on this foundation that appears on major platforms like Android and Chrome.
This chart ranks AI models by Chatbot Arena ELO score. A high score (highest number) indicates the size of the user’s preference. The Gemma 3N is highly ranked in both popular proprietary and open models.
Gemma 3N utilizes Google Deepmind Innovation, known as Layer-by-Layer Embedment (PLE), which significantly reduces RAM usage. The raw parameter counts are 5B and 8B, but this innovation allows you to run larger models on live streams from mobile devices or from clouds. This allows the model to operate with dynamic memory footprints of 2GB and 3GB, using memory overhead comparable to 2B and 4B models. See the documentation for more information.
By exploring the Gemma 3N, developers can get an initial preview of Open Model’s core features and mobile-first architecture innovations available on Android and Chrome with Gemini Nano.
In this post, we’ll explain new features of Gemma 3N, its approach to responsible development, and how to access today’s preview.
Key Features of Gemma 3n
Engineer the fast, low footprint AI experience introduced by Gemma 3N.
Optimized on-device performance and efficiency: Gemma 3N starts responding about 1.5 times faster with mobile with reduced memory footprint achieved through Mobile (compared to Gemma 3 4B) and innovations such as per-layer embedding, KVC sharing, and advanced activation quantization. Submodel (thanks to Matformer training). This gives you the flexibility to dynamically trade off performance and quality on the fly without hosting individual models. Additionally, we introduce the Mix’n’match feature in the Gemma 3N to dynamically create submodels from 4B models that can best fit the quality/latency trade-offs associated with specific use cases. Stay tuned for more details on this investigation in future technical reports. Privacy-First and Offline Support: Local Run enables features that ensure that users respect their privacy and functionality, even without an internet connection. Multimodal Understanding Enhanced with Audio: GEMMA 3N can greatly enhance audio, text and images, and greatly enhance video understanding. Its audio capabilities allow the model to perform high quality automatic speech recognition (transcription) and translation (speech to translated text). Furthermore, the model accepts interleaved inputs across modalities, allowing for understanding complex multimodal interactions. (Coming soon) Improved multilingual functionality: Improved multilingual performance, especially in Japanese, German, Korean, Spanish and French. Multilingual benchmarks such as 50.1% for WMT24++ (CHRF) reflect strong performance.
The MMLU performance of this chart show and the model size of the Gemma 3N’s Mix-n-Match (preprocessing) feature.
Unlock new on-the-go experiences
Gemma 3N empowers a new wave of intelligent on-the-go applications by enabling developers to:
Understand real-time visual and auditory cues from your user’s environment and build a live, interactive experience that supports.
2. The power of deeper understanding and contextual text generation using all personalized devices, combining audio, image, video and text input.
3. Develop advanced audio-centric applications, including real-time speech transcription, translation, and rich audio-driven interactions.
Here are some overviews and the types of experiences you can build:
We build together responsibly
Our commitment to responsible AI development is of paramount importance. Gemma 3N, like all Gemma models, received rigorous safety assessments, data governance and fine-tuning adjustments with safety policies. Approach open models with careful risk assessments and continually improve practices as AI landscapes evolve.
Let’s get started: Today we’ll preview Gemma 3N
I’m excited to get Gemma 3N through previews starting today.
Initial access (currently available):
Cloud-based Exploration with Google AI Studio: Try Gemma 3N directly in your Google AI Studio browser. No setup is required. Developing devices with Google AI Edge: For developers looking to integrate Gemma 3N locally, Google AI Edge provides tools and libraries. Today you can start the ability to understand/generate text and images.
Gemma 3N marks the next step in democratizing access to cutting-edge, efficient AI. Starting with today’s preview, we’re extremely excited to see what you’ll build as we gradually make this technology available.
Check out this announcement and all Google I/O 2025 updates on io.google starting May 22nd.
