What’s new in Gemma 4?

By Google DeepMind

Categories: AI, Product

Summary

Google DeepMind released Gemma 4 under Apache 2.0 open-source license with models ranging from 2B to 31B parameters, featuring native tool use, 250K token context windows, and support for 140+ languages—enabling developers to run frontier-grade AI directly on local hardware without data leaving their control.

Key Takeaways

1. 26B mixture-of-experts model with only 3.8B activated parameters delivers frontier intelligence with exceptional speed, optimizing token efficiency for resource-constrained environments while maintaining reasoning quality.
2. Gemma 4 supports 250K token context window enabling analysis of entire codebases and multi-turn agentic workflows in a single prompt—critical capability for enterprise code review and documentation use cases.
3. Native tool-use support built into the model architecture allows developers to build autonomous agents that plan and execute actions without custom prompt engineering or external orchestration frameworks.
4. 2B and 4B effective models with combined audio and vision capabilities enable real-time multimodal processing on mobile and IoT devices across 140+ languages—expanding edge AI deployments beyond text.
5. Apache 2.0 open-source licensing with enterprise-grade security protocols (matching proprietary models) removes licensing friction for builders and enterprises while maintaining production-ready safety standards.

Topics

• Open-Source LLM Licensing
• Local Model Inference
• Agentic AI Architecture
• Multimodal Edge Computing
• Mixture of Experts Optimization

Transcript Excerpt

We are thrilled to announce Jimma [music] 4. Built from the same world-class research and technology behind Gemini 3. Jima 4 is a family of open models designed to run directly on the hardware you own, phone, laptops, [music] and desktop. For the first time ever, we are releasing Gemma under an open-source Apache 2.0 license. Jimma 4 is built for the agentic era. It can handle complex logic, multi-step [music] planning, and agentic workflows, making optimal use of tokens for its intelligence. Th...