Google DeepMind has launched Gemma 4, a family of four open-weight AI models licensed under Apache 2.0, marking a major escalation in the U.S. open-source AI race.
The lineup includes the 31B Dense model-ranking third globally on Arena AI’s leaderboard-and a 26B Mixture of Experts model at sixth place. Both outperform models 20 times their size, according to Google’s benchmarks.

The smaller E2B and E4B variants are optimized for edge devices like smartphones and Raspberry Pis, with native audio support, 128K context windows, and zero-latency offline operation. The larger models support 256K context, structured JSON output, and native image/video processing-fitting on a single 80GB NVIDIA H100.
Unlike prior Gemma releases, Apache 2.0 removes legal barriers, enabling commercial use, modification, and redistribution without restrictions.
Google claims Gemma 4 is the best open model for its size class. Independent testing confirms strong code-generation reliability, though creative outputs remain serviceable rather than inspired.

This release directly challenges Chinese-dominated open models like Qwen and DeepSeek, reasserting U.S. leadership in locally deployable AI infrastructure.