Gemma 3 on mobile and web with Google AI Edge- Google Developers Blog

Introduces Gemma 3 1B, a new model size for on-device SLMs on mobile and web. Details performance metrics (2585 tok/sec on prefill) and use cases like data captioning, in-game dialog, smart reply, and document Q&A with an AI Edge RAG SDK. Provides a demo app for Android and instructions for loading the model from Hugging Face. Explains under-the-hood optimizations including 4-bit quantization, optimized KV cache layouts (improving latency by ~25% on CPU, ~20% on GPU), cached optimized tensor layouts, and GPU weight sharing for LLM inference. Mentions plans for supporting more third-party models and further memory reduction in 2025.