Google DeepMind Debuts Gemma 4: The New Standard for Efficient Open AI Models

Google DeepMind has officially launched Gemma 4, its most intelligent family of open models to date. Built on the same research foundation as the proprietary Gemini 3, these models are designed to deliver an unprecedented ratio of intelligence to parameter count. The entire lineup is available under the commercially permissive Apache 2.0 license, providing developers with the freedom to build and deploy without restrictive barriers. The family includes four distinct sizes tailored for various hardware environments. The E2B and E4B models are engineered for edge devices, while the 26B Mixture-of-Experts and 31B Dense models target high-performance workstations. Notably, the 31B model currently ranks as the third-highest open model on the Arena AI leaderboard, outperforming systems many times its physical size.

Beyond text generation, Gemma 4 introduces native multimodality across the board. Every model can process images and video, while the edge-optimized versions also feature native audio input for real-time speech recognition. This allows sophisticated visual tasks like OCR and chart understanding to be performed locally on mobile phones or IoT hardware.

To support autonomous agents, Google integrated native support for function-calling and structured JSON output. Developers can now build workflows where the AI interacts directly with external APIs with high reliability. A new thinking mode also allows models to display their internal logic before providing a final answer, improving transparency in complex reasoning.

Global accessibility is a cornerstone of this release, with native training spanning more than 140 languages. The models support expanded context windows, with larger versions handling up to 256K tokens. This capacity enables the processing of entire code repositories or long documents without losing coherence.

Optimization for NVIDIA, Qualcomm, and MediaTek hardware ensures that frontier-class AI runs efficiently on everything from a Raspberry Pi to a data center GPU. By prioritizing performance, Google aims to empower a global community of innovators to push the boundaries of open AI.