Google DeepMind releases Gemma 4, claiming state-of-the-art performance across four model sizes

Google DeepMind announced Gemma 4, a family of open-source language models released under an Apache 2.0 license. The lineup comprises four distinct sizes: a 2-parameter effective model, a 4-parameter effective model, a 26-parameter Mixture of Experts variant, and a 31-parameter dense model. The company positions these as purpose-built for advanced reasoning and autonomous agent workflows, emphasizing efficiency across hardware targets from mobile devices to workstation accelerators.

According to the announcement, Gemma 4's 31B model currently ranks third on Arena.ai's chat leaderboard as of April 1, 2026, with the 26B variant ranking sixth. The company claims these models outperform certain larger competitors on benchmarks. All variants feature context windows ranging from 128,000 tokens in smaller edge models to 256,000 tokens in larger versions, and natively process video, images, and audio input across over 140 languages.

The models support function-calling, structured JSON output, and system-level instruction handling, capabilities intended to enable autonomous agent construction. The announcement highlights prior community adoption, citing over 400 million downloads of the original Gemma line and more than 100,000 community variants built from those models. Specific prior use cases mentioned include collaboration with Yale University on molecular research and development of a Bulgarian-language model.