Google DeepMind releases Gemma 4: Frontier multimodal intelligence on device

AI Impact Summary

Google DeepMind is releasing Gemma 4, a new family of multimodal models designed for on-device intelligence. These models, available on Hugging Face, boast Apache 2 licenses, high quality scores, and support for image, text, and audio inputs, making them suitable for agents and inference engines. The architecture incorporates innovations like Per-Layer Embeddings and a Shared KV Cache to optimize performance and memory usage, particularly for long context and quantization.