What’s new in Gemma 4

Google DeepMind has announced Gemma 4, a new family of open models now released under the Apache 2.0 license. The family includes a 26B Mixture of Experts model (3.8B activated parameters), a 31B Dense model, and efficient 2B and 4B models targeting mobile and IoT devices. Key features include a 250K token context window, native tool use for agentic workflows, combined audio and vision support, and multilingual support for over 140 languages. The models are built for local execution on personal hardware — phones, laptops, and desktops — without requiring data to leave the user's environment. Enterprise security protocols from Google DeepMind are applied to all models.