Google DeepMind Unleashes Gemma 4 Family – #3 Open Model Globally Overnight!

Google DeepMind has significantly expanded the open-source AI frontier with the release of its Gemma 4 model family. The release includes four distinct models: the smaller E2B and E4B, a 26B parameter mixture-of-experts (MoE) model, and a flagship 31B parameter dense model. Built on the same foundational technology as the proprietary Gemini 3, the family is designed for advanced reasoning and enabling agentic AI workflows—where AI can autonomously take actions. Crucially, all models are released under the permissive Apache 2.0 license, allowing for widespread commercial and research use.

Benchmark performance was immediate and impressive. The 31B dense model shot to the #3 position on the prestigious Chatbot Arena leaderboard for open models globally, with the 26B MoE model securing the #6 spot. Early testers, like developer Simon Willison, noted a clear quality progression across model sizes, even in creative tasks like drawing pelicans. NVIDIA confirmed day-one optimization for its ecosystem, including consumer RTX GPUs, enterprise DGX Spark systems, and Jetson edge devices, ensuring strong performance out of the gate. While some early quantization files (GGUF) for the 31B model had issues, the release marks a major step in providing state-of-the-art, accessible intelligence for developers and enterprises looking to build beyond closed APIs.