Gemma 4 Released

The lineup consists of 4 models — 31B Dense, 26B-A4B MoE, E4B, and E2B for smartphones (the architecture of these two is similar to Gemma 3n). All of them are reasoning models and can handle images with videos, while E4B and E2B can also process audio. Context is supported up to 256k for the two larger models and 128k for the two smaller ones. All models, except for the smallest one, outperform Gemma 3 27B on all presented benchmarks (the smallest one still lags behind 27B on a couple of benchmarks). All weights are published under Apache 2.0, and the base models are also available. There is currently no technical report.