Google DeepMind Debuts Gemma 3n, Bringing Advanced Multimodal AI On-Device

Distilled summaries of the latest news shaping the future of AI, Web3 and Quantum

May 23, 2025

Details

  • Gemma 3n introduces Per-Layer Embeddings (PLE), reducing RAM usage nearly threefold and making it possible to run complex AI directly on mobile devices.
  • The 5B and 8B parameter models are optimized to perform like much smaller 2B or 4B models, needing only 2GB or 3GB of memory for mobile or cloud implementation.
  • The system supports multimodal input—including text, images at up to 768x768 resolution, audio at 6.25 tokens per second, and up to 32,000 token context windows.
  • Its MatFormer architecture and conditional parameter loading allow the system to selectively activate necessary modules, efficiently managing device resources.
  • Gemma 3n is now available in early preview via Google AI Studio, launched during Google I/O 2025 as part of a push for mobile-first AI.

Impact

By slashing memory needs and supporting multimodal processing, Gemma 3n enables advanced AI functions to run smoothly on everyday smartphones. This technical leap helps drive the industry closer to seamless, privacy-focused on-device AI, challenging rivals to prioritize efficiency and device compatibility. Gemma 3n’s architecture could shape future innovations in mobile and embedded AI.

Read more