Gemma 4 Ships Under Apache 2.0: 31B Dense Model Beats 400B Rivals — AIME Math 20.8% → 89.2%, LiveCodeBench 29.1% → 80.0%

Google Blog·high signal

Google released Gemma 4 on April 2 with four variants (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0 — the first fully permissive license in the Gemma family, removing all prior commercial restrictions. Performance gains over Gemma 3 are massive: AIME math jumps from 20.8% to 89.2%, coding from 29.1% to 80.0%, science from 42.4% to 84.3%. All variants support native text/vision/audio, 256K context, and 140+ languages. The 31B dense model runs on a single consumer GPU. For builders: the best open model you can legally ship in any product, with frontier-class reasoning.