Google DeepMind has unveiled Gemma 3n, a cutting-edge open model designed for fast, multimodal AI directly on mobile devices. The new architecture enables lightning-fast performance with significantly reduced memory footprint through innovations like Per-Layer Embeddings, allowing the 5B/8B parameter models to operate with just 2GB/3GB of RAM.

Developed in partnership with Qualcomm Technologies, MediaTek, and Samsung's System LSI, Gemma 3n features a unique 2-in-1 model that includes a nested 2B submodel within a 4B footprint. This provides developers flexibility to dynamically trade performance and quality without hosting separate models. The model also expands multimodal understanding to include audio processing for transcription and translation.

The preview of Gemma 3n is now available through Google AI Studio for cloud-based exploration and Google AI Edge for on-device development. This release marks Google's continued push toward democratizing access to efficient AI technology that powers next-generation applications while maintaining privacy through local execution.