Google DeepMind Releases Gemma 4: Open AI Models for Consumer Devices

Google DeepMind has launched Gemma 4, a new family of open-source AI models designed to deliver high-level intelligence and agentic capabilities directly to consumer hardware, including mobile phones and local workstations. Built on the same foundational research as Gemini 3, the Gemma 4 family includes small sizes for edge deployment, a dense 31-billion parameter model, and a 26-billion parameter Mixture-of-Experts architecture. The release features an Apache 2 license, granting developers full flexibility to commercialize applications. Since its launch last month, Gemma 4 has achieved over 50 million downloads. In a recent discussion, Omar Sanseviero from Google DeepMind highlighted the model's efficiency, demonstrating offline agentic workflows such as generating SVGs and executing complex tasks on Android devices. Demonstrations included a local food tour agent that performed multi-step reasoning for budget-friendly dining in Seattle, and an autonomous coding example where the model wrote Python code to simulate physics, self-correcting execution errors to successfully animate a bouncing ball. This release marks a significant shift towards running powerful AI agents with minimal hardware overhead.