Announcing Gemma 3n preview: powerful, efficient, mobile-first AI

After the relevant GEMMA 3 and Gemma 3 Qat launches, a family of open models on the latest models capable of running one cloud or desktop, we pay our vision to reach artificial intelligence. GEMMA 3 provided powerful abilities to developers, and we are now expanding that vision to AI in actual time working directly on the devices you use every day-your phones, tablets and laptops.

To operate the next generation of artificial intelligence on the device and support a variety of applications, including the introduction of Gemini Nano capabilities, we have designed a new advanced structure. This next generation Foundation was created in close cooperation with mobile drivers such as Qualcomm Technologies, Mediaatek and Samsung’s System LSI, and it was improved in favor of rapid artificial, multimedia intelligence, providing personal and truly special experiences on your device.

GEMMA 3N is our first open model based on this leading joint architecture, allowing developers to start experimenting with this technology today in an early preview. The same advanced architecture also operates the next generation of Gemini Nano, which brings these capabilities to a wide range of features in Google applications and our ecosystem on devices, and will be available later this year. GEMMA 3N enables you to start building on this basis, which will reach major platforms such as Android and Chrome.

GEMMA 3N takes advantage of the Google DeepMind creation called the content of each layer (PL) that provides a significant decrease in the use of RAM. Although the number of raw parameters is 5B and 8B, this innovation allows you to operate larger models on mobile devices or live broadcasts of the cloud, with a memory in the sky of the 2B and 4B model, which means that models can work with a dynamic memory imprint of 2 GB and only 3 GB. Learn more in our documents.

By exploring GEMMA 3N, developers can get an early preview of the capabilities of the primary open model and the first architectural innovations of mobile phones that will be available on Android and Chrome with Gemini Nano.

In this post, we will explore the new gemma 3N capabilities, our approach to responsible development, and how you can access the inspection today.

The main capabilities of GEMMA 3N

GEMMA 3N, designed for low -horizon AI experiences.

• Improving performance and efficiency on devices: GEMMA 3N begins to respond about 1.5X faster on the mobile phone with a significant quality (compared to GEMMA 3 4B) and with a discounted memory stroke achieved through innovations such as layer implications, KVC sharing, and advanced activation measurement.

• Many flexibility: 1: A model has an active memory fingerprint 4B that automatically includes a sub -model of the 2B active memory (thanks to the MATFORMER). This provides flexibility to exchange performance and quality dynamically without hosting separate models. We also offer the possibility of Mix’n’msh in GEMMA 3N to create 4B sub -models that can fit with the optimal defined use status – and the associated quality/cumin. Stay tuned for this research in our next technical report.

• The first privacy and not connected to the Internet: The local implementation enables the features that respect the user’s privacy and its functions reliably, even without an internet connection.

• Excluded multimedia understanding with sound: GEMMA 3N can understand, process sound, text and images, and advances a greatly enhanced video. Its vocal capabilities allow the model to perform high -quality automatic speech recognition (copy) and translation (speech to translated text). In addition, the model accepts the interlocking inputs across the methods, allowing the understanding of the complex multimedia reactions. (General Implementation soon)

• Multi -language potential improvement: Improving multi -language performance, especially in Japanese, German, Korean, Spanish and French. Strong performance is reflected on multi -language standards such as 50.1 % on WMT24 ++ (Chrf).

Opening new experiences while moving

GEMMA 3N will enable a new wave of smart applications while enabling developers to:

1. Build live interactive experiences This understands and responds to visual and auditory signals in the actual time of the user environment.

2. A deeper power of understanding The generation of contextual texts using audio, image, video and common text inputs-all of this was addressed at a special level.

3. Development of advanced applications about soundIncluding copying speech in real time, translation, and rich sound reactions.

Below is an overview and types of experiments that you can build:

Building responsibility, together

Our commitment to developing responsible artificial intelligence is crucial. GEMMA 3N, like all GEMMA models, underwent strict safety assessments, data governance, and align their control with our safety policies. We are dealing with open models with an accurate rating of risk, and we constantly improve our practices with the development of the male scene.

Start: Preview GEMMA 3N today

We are excited to get GEMMA 3N in your hands with a preview starting today:

Initial access (now available):

• Document to the counterpart group with Google Ai Studio: Try GEMMA 3N directly in your Google Ai Studio – Don’t Prepare. Explore the abilities to enter the text immediately.

• Development of the device with Google Ai Edge: For developers looking to integrate GEMMA 3N, Google Ai Edge provides tools and libraries. You can start the possibilities of understanding text and images/obstetrics today.

GEMMA 3N is the next step in the democratic character to reach advanced and effective artificial intelligence. We are incredibly enthusiastic about what you will build as we make this technology gradually available, starting with today’s preview.

Explore this advertisement and all Google I/O 2025 updates on IO.GOOGLE starting from May 22.