Google Releases Gemma 4 Open Models With Apache 2.0 License and Native Agentic Capabilities
Google released Gemma 4 on April 2, 2026, marking a significant shift in its open-weight model strategy with a move to the permissive Apache 2.0 license and purpose-built support for agentic AI workflows. The release includes four model variants—Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense—all available via Google AI Studio and Hugging Face.
Apache 2.0 Licensing Enables Unrestricted Commercial Deployment
The shift to Apache 2.0 licensing represents a major change from previous Gemma versions, which carried more restrictive terms. VentureBeat noted that "that license change may matter more than benchmarks," as Apache 2.0 puts Gemma 4 on par with Llama and other truly open models for commercial deployment without Google's prior restrictions. This licensing change enables developers to use Gemma 4 in production environments without seeking Google's approval.
Models Feature Native Multimodal and Agentic Workflow Support
All Gemma 4 models natively process video and images with variable resolution support, exceling at visual tasks including OCR and chart understanding. The E2B and E4B edge models feature native audio input for speech recognition and understanding, while offering 128K context windows. The larger 26B and 31B models provide up to 256K context windows.
Google explicitly positioned Gemma 4 as purpose-built for advanced reasoning and agentic workflows. The models include native support for function-calling, structured JSON output, and native system instructions—capabilities that enable building autonomous agents that interact with tools and APIs without complex prompting hacks or wrapper libraries. The 31B Dense model ranks #3 on the Arena AI text leaderboard globally among all open models, while the 26B MoE model ranks #6.
26B MoE Model Optimizes for Latency With Sparse Activation
The 26B MoE model focuses on latency optimization by activating only 3.8B of its total parameters during inference, delivering exceptionally fast token generation. This sparse activation approach contrasts with the 31B Dense model, which maximizes raw quality by using all parameters. The combination gives developers flexibility to choose between speed and maximum performance based on their deployment requirements.
Key Takeaways
- Google released Gemma 4 on April 2, 2026, with Apache 2.0 licensing that enables unrestricted commercial use, removing previous restrictive terms
- Four model variants range from 2B to 31B parameters, with the 31B Dense ranking #3 globally on the Arena AI text leaderboard
- All models feature native multimodal support for video, images, and audio (E2B/E4B only), with context windows up to 256K tokens
- Purpose-built agentic capabilities include native function-calling, structured JSON output, and system instructions for production agent deployment
- The 26B MoE model activates only 3.8B parameters during inference for optimized latency while maintaining competitive performance