Gemma 4 is an open-source AI model from Google DeepMind designed to run locally on personal devices, offering advanced capabilities like handling up to 250,000 tokens, multi-step planning, and native tool use for autonomous agent workflows. It comes in various sizes optimized for different hardware, supports over 140 languages, prioritizes security, and empowers users with powerful, privacy-focused AI for diverse applications.
Gemma 4 is the latest release in the Gemma family of open models, built on the advanced research and technology behind Gemini 3. Designed to run directly on personal hardware such as phones, laptops, and desktops, Gemma 4 is now available under an open-source Apache 2.0 license. This release marks a significant step towards empowering users with powerful AI models that operate locally, ensuring greater control and privacy.
The model is tailored for the agentic era, capable of handling complex logic, multi-step planning, and agentic workflows efficiently. One of its standout features is the ability to work with a context window of up to 250,000 tokens, enabling it to analyze entire codebases and support multi-turn agentic use cases. Additionally, Gemma 4 includes native support for tool use, allowing users to build intelligent agents that can plan and act autonomously on their behalf.
Gemma 4 comes in several variants to suit different needs. The 26 billion parameter mixture of experts (MoE) model and the 31 billion parameter dense model offer frontier-level intelligence that can run locally on personal computers. The 26B MoE model is optimized for speed with 3.8 billion activated parameters, while the 31B dense model focuses on output quality. These models enable state-of-the-art reasoning and coding workflows without requiring data to leave the user’s environment.
For mobile and IoT devices, Gemma 4 offers effective 2 billion and 4 billion parameter models engineered for maximum memory efficiency. These smaller models support combined audio and vision processing in real-time, allowing them to perceive and interpret the world around them. They also natively support over 140 languages, making them versatile for multilingual and agentic tasks across diverse applications.
Security is a top priority for Gemma 4, which is developed by Google DeepMind and adheres to the same rigorous security protocols as proprietary models. This ensures a trusted foundation for enterprises and developers to build upon. Users can download the model weights and start experimenting immediately, integrating Gemma 4 with familiar tools to create innovative AI-powered solutions. The release invites the community to explore and expand the possibilities of open, locally run AI models.