Gemma 4: compact, open models tuned for advanced reasoning and edge use

Gemma 4 is a new family of open models designed for advanced reasoning and agentic workflows, delivering high intelligence-per-parameter and released under an Apache 2.0 license. Built from the same research foundation as Gemini 3, the models complement proprietary offerings and follow strong community uptake—developers have downloaded Gemma over 400 million times, producing more than 100,000 variants.

The family ships in four sizes—Effective 2B (E2B), Effective 4B (E4B), 26B Mixture of Experts (MoE) and 31B Dense—and moves beyond simple chat to handle multi-step planning, deep logic and agentic flows. On Arena AI’s text leaderboard as of 4/1, the 31B model ranks #3 among open models and the 26B ranks #6, outperforming models many times larger.

Edge-oriented E2B and E4B prioritize multimodal capability, low latency and on-device use.

gemma 4, open models, apache 2.0, gemini 3, effective 2b, effective 4b, 26b moe, 31b dense, agentic workflows, arena ai