Gemma 4 Released: The Most Capable Apache 2.0 AI Models

4/3/2026

The fundamental constraints of artificial intelligence development have been definitively removed. Building upon a massive community momentum of over 400 million downloads, Google has officially unveiled Gemma 4—its most intelligent open model family to date. Stripping away restrictive barriers, the entire architecture is released under a commercially permissive Apache 2.0 license. Purpose-built for advanced reasoning and complex agentic workflows, Gemma 4 is delivered in four highly versatile sizes: Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE), and 31B Dense. https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/gemma-4__elo-score__eval__dark_Web.png The performance data redefines the boundaries of efficiency. On the industry-standard Arena AI text leaderboard, the 31B model dominates by securing the #3 spot globally among open models with an Elo score of 1452. Concurrently, the 26B model achieves an Elo score of 1441, taking the #6 position and decisively outcompeting alternative models that are 20 times its size. Natively trained on over 140 languages, the series natively supports autonomous function-calling, structured JSON outputs, and high-quality offline code generation. Furthermore, the E2B and E4B edge models feature a 128K context window, while the larger variants extend up to 256K, seamlessly processing massive repositories in a single prompt. Hardware efficiency operates at the very core of Gemma 4. The unquantized bfloat16 weights of the 26B and 31B models fit efficiently onto a single 80GB NVIDIA H100 GPU. The 26B MoE variant is hyper-optimized for latency, activating precisely 3.8 billion parameters during inference to maximize tokens-per-second output. At the computing edge, the multimodal E2B and E4B models completely redefine on-device utility. They process audio and vision natively, running completely offline with near-zero latency across smartphones, Raspberry Pi, and NVIDIA Jetson Orin Nano. This allows Android developers to immediately prototype agentic flows in the AICore Developer Preview, ensuring strict forward-compatibility with Gemini Nano 4. https://storage.googleapis.com/gweb-uniblog-publish-prod/documents/gemma-4-table_light_Web_with_Arena.jpg