Google Launches Gemma 3n Open-Source AI Model for Deployment

Google has unveiled the full version of Gemma 3n, its latest open-source artificial intelligence model, designed for on-device applications. Announced in May, this innovative model boasts several architectural enhancements and can operate with just 2GB of RAM, making it suitable for deployment on smartphones equipped with AI processing capabilities. The release marks a significant addition to the Gemma 3 family, which includes previous models like Gemma 3 and GemmaSign.

Gemma 3n: A Multimodal AI Model

In a recent blog post, Google detailed the features of Gemma 3n, emphasizing its multimodal capabilities. This model can process various input types, including images, audio, video, and text, although it is limited to generating text outputs. Notably, Gemma 3n supports 140 languages for text and 35 languages for multimodal inputs, enhancing its accessibility for a global audience. As an open-source model, Google has made the model weights and a comprehensive cookbook available to the community under a permissive Gemma license, allowing for both academic and commercial applications.

The architecture of Gemma 3n is built on the Matryoshka Transformer, or MatFormer, which employs a nested design reminiscent of Russian nesting dolls. This innovative structure enables the training of AI models with varying parameter sizes, optimizing performance for mobile devices. Gemma 3n is offered in two sizes—E2B and E4B—representing effective parameters. While the model sizes are five billion and eight billion parameters, respectively, the active parameters are reduced to two and four billion, thanks to a technique called Per-Layer Embeddings (PLE). This approach allows only essential parameters to be loaded into fast memory, while others remain accessible through additional layers.

Flexible Model Sizes and Customization

The MatFormer architecture allows the E4B variant to nest the E2B model, enabling simultaneous training of both models. This design provides users with the flexibility to choose between the E4B for advanced operations and the E2B for quicker outputs, all while maintaining consistent quality in processing and results. Furthermore, Google is empowering users to create custom-sized models by adjusting specific internal components. To facilitate this, the company is introducing the MatFormer Lab tool, which enables developers to experiment with various configurations to identify optimal model sizes.

Availability and Access

Gemma 3n is currently accessible for download through Google’s listings on Hugging Face and Kaggle. Additionally, users can explore the model via Google AI Studio, where they can experiment with its capabilities. Notably, the Gemma models can also be deployed directly to Cloud Run from AI Studio, streamlining the integration process for developers and researchers. This release underscores Google’s commitment to advancing AI technology while making it more accessible to a wider audience.

Observer Voice is the one stop site for National, International news, Sports, Editor’s Choice, Art/culture contents, Quotes and much more. We also cover historical contents. Historical contents includes World History, Indian History, and what happened today. The website also covers Entertainment across the India and World.

Follow Us on Twitter, Instagram, Facebook, & LinkedIn