Welcome Gemma 4: Frontier multimodal intelligence on device

Google DeepMind's Gemma 4 family launches on Hugging Face with Apache 2 licenses, multimodal support (image, text, audio), and sizes ranging from 2B to 27B parameters with up to 128K context windows. Models are available across transformers, llama.cpp, MLX, WebGPU, and Rust, with on-device deployment as a first-class target. Hugging Face reports the models score at the Pareto frontier on arena benchmarks and are strong enough out of the box to make fine-tuning examples hard to find.