Introducing Gemma 3n: The developer guide- Google Developers Blog

Introduces Gemma 3n, a mobile-first multimodal LLM architecture optimized for on-device deployment. Key innovations include the MatFormer architecture for elastic inference (allowing pre-extracted E2B/E4B models and custom sizes via Mix-n-Match), Per-Layer Embeddings (PLE) for memory efficiency by offloading embeddings to CPU, and KV Cache Sharing for faster long-context processing. Gemma 3n integrates an advanced audio encoder (USM) for ASR and AST, and a new MobileNet-V5-300M vision encoder for high-throughput multimodal tasks on edge devices. The post details benchmark results and community support across various tools.