Welcome Gemma 4: Frontier multimodal intelligence on device

Google DeepMind released Gemma 4, a family of multimodal models now available on Hugging Face under Apache 2 licenses. The models support image, text, and audio inputs across five sizes, with the 31B dense variant achieving an estimated LMArena score of 1452. Gemma 4 introduces Per-Layer Embeddings and a shared KV cache for efficiency. Hugging Face contributors including merve, Pedro Cuenca, and Sergio Paniego detailed integration with transformers, llama.cpp, MLX, and other frameworks. DiffusionGemma enables text generation via diffusion.