Google announces Gemma 4 open AI models, switches to Apache 2.0 license

Google opened Gemma 4 to developers starting today, offering four sizes optimized for local use.

The family includes a 26B Mixture of Experts (only 3.8B parameters active in inference) and a 31B Dense focused on quality and expected to be fine-tuned.

Smaller E2B and E4B models run as effective 2B or 4B variants for devices like smartphones, Raspberry Pi, and Jetson Nano.

Google says Gemma 4 uses less memory and battery than Gemma 3, delivers near-zero latency, and is the most capable model for local hardware.

Google abandoned the custom Gemma license, and Gemma 31B will debut at #3 on the Arena list behind GLM-5 and Kimi 2.5.