Google Gemma 4 ships under Apache 2.0 — four sizes, MoE, multimodal, 256K context
Gemma 4 (April 2) arrives in E2B / E4B / 26B MoE / 31B Dense variants with native image+video everywhere and native audio on the smaller models. 256K context, 140+ languages, agentic-workflow-oriented. The 31B Dense reportedly hit #3 on Arena's text leaderboard.
The license is the most notable change from Gemma 3: Google moved Gemma 4 to Apache 2.0, making it materially more usable for commercial and derivative work than the prior Gemma terms allowed.
Model lineup is unusual — two "Effective" sizes (E2B and E4B) for local/on-device, a 26B Mixture-of-Experts for cloud, and a 31B Dense flagship. The on-device story is interesting: with reduced RAM requirements vs. Gemma 3 and native audio input, Google is positioning Gemma 4 as the default local agentic model on Android (per the AICore developer preview).
Benchmark claims worth tracking independently: Google reports the 31B Dense at #3 on Arena's text leaderboard and the 26B MoE at #6 — high if confirmed by independent eval. 256K context is the table-stakes number for serious agentic work.