Welcome Gemma 4: Frontier multimodal intelligence on device

Google DeepMind has released the Gemma 4 family of open, multimodal models on Hugging Face, supporting image, text, and audio inputs. These models are designed for high performance and on-device use, coming in four sizes including dense and mixture-of-experts (MoE) variants. Key architectural features include alternating attention layers, Per-Layer Embeddings (PLE), and a Shared KV Cache to improve efficiency and support long contexts. The release includes broad support for popular libraries like transformers, llama.cpp, and MLX, facilitating widespread deployment and fine-tuning.