What’s new in Gemma 4

Key Concepts

• Gemma 4: A new family of open-weights models from Google DeepMind.
• Agentic Workflows: AI systems capable of complex logic, multi-step planning, and tool usage.
• Apache 2.0 License: An open-source license allowing for broad commercial and personal use.
• Mixture of Experts (MoE): A model architecture that activates only a subset of parameters per token to increase efficiency.
• Context Window: The amount of data (tokens) a model can process at once (up to 256k tokens).
• Edge AI: Running models locally on hardware like phones, laptops, and IoT devices.

---

Introduction to Gemma 4

Olivier, a Group Product Manager on the Gemma team, announced the launch of Gemma 4. Building on the success of previous iterations—which saw over 400 million downloads and 100,000 variants—Gemma 4 is designed specifically for the "agentic era." For the first time, the model family is released under an Apache 2.0 license, providing developers with greater freedom for integration and innovation.

Model Architecture and Capabilities

Gemma 4 is engineered to run locally on personal hardware, ensuring data privacy by keeping information within the user's controlled environment.

• Agentic Intelligence: The models are optimized for complex logic, multi-step planning, and native tool use, allowing them to act on behalf of the user.
• Context Window: The larger models support a context window of up to 256,000 tokens, enabling the analysis of entire codebases and long-form, multi-turn agentic interactions.
• Multilingual Support: The entire family natively supports over 140 languages.

The Gemma 4 Model Family

The lineup is categorized by size and use case:

1. High-Performance Models (26B MoE & 31B Dense):
   • 26B Mixture of Experts (MoE): Features 3.8 billion activated parameters. It is designed for exceptional speed while maintaining high-level reasoning.
   • 31B Dense Model: Optimized specifically for maximum output quality, providing "frontier intelligence" on personal computers.
2. Efficiency-Focused Models (2B & 4B):
   • Engineered for maximum memory efficiency.
   • Designed for mobile and IoT devices.
   • Multimodal Capabilities: These models include native audio and vision support, allowing for real-time processing of sensory data.

Security and Trust

A core pillar of the Gemma 4 release is enterprise-grade security. Developed by Google DeepMind, these models undergo the same rigorous security protocols as Google’s proprietary models. This provides a "trusted foundation" for developers and enterprises to integrate Gemma 4 into their infrastructure without compromising safety.

Practical Application and Demonstration

The presentation highlighted the practical utility of the 2B model through a multilingual agentic task. The model successfully processed a request in a foreign language (Italian) and provided a coherent, accurate response in English, demonstrating its ability to handle real-time, cross-lingual agentic workflows.

Synthesis and Conclusion

Gemma 4 represents a significant shift toward local, agentic AI. By combining high-performance architectures (MoE and Dense) with extreme efficiency for edge devices (2B/4B), Google is enabling developers to build sophisticated, private, and secure AI agents. The transition to an Apache 2.0 license further cements Gemma 4 as a primary tool for the next generation of open-source AI development, moving beyond simple text generation into the realm of autonomous, tool-using agents.