Gemma 4 arrives under Apache 2.0 in four model sizes

The family spans edge models, a mixture-of-experts model and a dense model.

Capvise MediaEvent 2 Apr 2026Published 15 Oct 20261 min read

Photo: Earlier or representative hardware photograph: Google Pixel and Pixel XL smartphones (30120254046). The pictured model differs from the product discussed in this report. Credit: Maurizio Pesce from Milan, Italia, “Google Pixel and Pixel XL smartphones (30120254046).jpg,” Wikimedia Commons, CC BY 2.0. Adapted by resizing and WebP re-encoding. · CC BY 2.0 · Resized for web delivery

Weights developers can host

Google released Gemma 4 on 2 April in four sizes: Effective 2B, Effective 4B, a 26B mixture-of-experts model and a 31B dense model. The family uses the commercially permissive Apache 2.0 licence, changing the practical permissions around adapting and deploying the weights. Google designed the smaller models for edge devices and the larger ones for more capable hardware.

The edge models support 128K context, compared with up to 256K for the larger variants. All process images and video; the E2B and E4B models also accept audio. Function calling and structured JSON support target applications beyond chat. Local deployment still requires capacity planning: open weights remove a hosted API dependency, but not memory, inference or maintenance costs.