Google unveils open-weight Gemma 4 for low-power devices

Google announced Gemma 4, an open-weight AI model family based on Gemini 3 architecture.

The family includes Effective 2B, Effective 4B, 26B MoE, and 31B Dense models for low-power device inference.

All models handle image and video inputs; E2B and E4B also support native audio for on-device speech understanding.

Gemma 4 supports native function calls and structured JSON output and is available under the Apache 2.0 license on Google Cloud and open-weight repositories.