Google DeepMind has fully released Gemma 3n, a mobile-first multimodal model designed for on-device applications. This new architecture supports image, audio, video, and text inputs, with text outputs, and is optimized for efficiency, offering versions with effective parameters of 2B and 4B that mimic the memory footprint of traditional 2B and 4B models. Gemma 3n introduces novel components like MatFormer for flexibility and Per Layer Embeddings for memory efficiency, achieving strong performance in multilinguality, math, coding, and reasoning, with the E4B version surpassing 1300 on the LMArena benchmark. The model is available through popular developer tools and integrates with the growing Gemmaverse ecosystem. AI
Summary written by None from 4 sources. How we write summaries →
RANK_REASON Google DeepMind released Gemma 3n, a new multimodal model with novel architecture and strong benchmark performance for on-device applications.
