Google DeepMind launches Gemma 4, its highest-performing family of open models

Google DeepMind launches Gemma 4, a new open model family in four sizes under Apache 2.0 license, supporting 140+ languages and up to 256K token contexts.

Google DeepMind releases Gemma 4 in four sizes, from lightweight mobile-optimized versions (E2B, E4B) to more powerful models (26B MoE, 31B Dense), with the latter ranking third among open models on Arena AI. The family supports over 140 languages, contexts up to 256K tokens, and covers code, vision, audio, and agentic workflows. All models are released under the Apache 2.0 license and are available on Hugging Face, Ollama, Google AI Studio, and several other environments.