This video introduces Gemma 3N, Google's latest open and accessible AI model designed for compelling on-device experiences and leveraging portable compute power. Gemma 3N shares the core technology with Gemini Nano models, offering developers an early preview of cutting-edge architecture tailored for efficiency and state-of-the-art on-device performance. Key innovations include superior quality for a small model, capabilities such as on-device function calling, interleaved text and images, and notably, the first in the Gemma family to understand audio and video inputs. The model also boasts an efficient inference profile optimized for mobile processors, resulting in significantly faster processing and lower memory usage. A unique 2-in-1 mat former architecture allows dynamic selection between peak quality and greater speed/lower resource usage within a single model's memory footprint, reinforcing Google's commitment to robust, real-world on-device AI.
Introducing Gemma 3N: Next-Generation On-Device AI
0:11 Introduction to Gemma 3N: Lucas, a product manager on the Gemma team, introduces Gemma 3N, a new model focused on unlocking compelling on-device experiences and leveraging portable compute.
0:36 Nano-Inspired Efficiency: The "N" in Gemma 3N highlights its direct lineage to nano-sized models, signifying a core design around efficiency and state-of-the-art on-device performance.
0:50 Gemini Nano Technology: Gemma 3N is built on the same technology as Gemini Nano models, offering developers an early preview of the architecture and capabilities powering next-generation on-device AI for platforms like Android and Chrome.
1:33 Best-Ever Small Model Quality: Gemma 3N delivers Google's best quality for a small model, meticulously tailored for on-device use cases.
1:46 Revolutionary Capabilities: It unlocks on-device function calling and interleaved text and images.
1:51 Multi-Modal Understanding: For the first time in the Gemma family, Gemma 3N is capable of understanding audio and video inputs, allowing natural user interaction with applications.
3:56 Optimized On-Device Performance: Gemma 3N features Google's most efficient inference profile for on-device AI, optimized for mobile processors, leading to significantly faster input processing (prefill) and less memory usage compared to previous models like Gemma 34B.
4:34 Unique 2-in-1 Architecture: The model includes a smaller embedded submodel as part of an innovative 2-in-1 mat former architecture, enabling dynamic choice between peak quality or greater speed and lower resource usage on the fly, all within the memory footprint of a single model.
5:01 Open and Accessible AI: Gemma 3N embodies Google's commitment to open and accessible AI, incorporating co-design insights from Android, Chrome, and Pixel teams for robust real-world on-device deployment.
5:21 Availability: Developers can explore the models using various tools and providers.