Gemma 4: Byte for byte, the most capable open models

Google DeepMind released Gemma 4 today, and the headline number is hard to ignore: a 31B-parameter model that ranks third on the Arena AI leaderboard among every open model out there, ahead of systems many times its size. Its sibling, a 26B mixture-of-experts model, lands at sixth. Google is calling this "intelligence-per-parameter," which is a fancy way of saying you get more brains without needing a server farm to run them.

The family comes in four flavors: E2B and E4B for phones and edge devices, plus the 26B MoE and 31B dense models for workstations and single H100 GPUs. That MoE model only activates 3.8 billion parameters per inference pass, which is the trick behind its speed — most of the network stays dormant for any given query. The 31B dense version, by contrast, uses everything it has, trading some latency for raw output quality, and Google positions it as the better base for fine-tuning.

What's notable is how far this pushes on-device AI. The E2B and E4B models were built with Qualcomm, MediaTek, and the Pixel team specifically to run offline on phones, Raspberry Pi boards, and Jetson Orin Nano chips, with near-zero latency and native audio input for speech tasks. Context windows scale up to 128K on the edge models and 256K on the bigger ones, enough to feed in a full codebase or a lengthy report in one shot. All four sizes handle images and video natively, and the whole family was trained across more than 140 languages, which is a serious commitment to markets outside English-speaking ones.

Google is also leaning hard into the licensing story here. Gemma 4 ships under Apache 2.0, a fully permissive open-source license with no usage restrictions, which the company frames explicitly around