Google launches Gemma 4 open multimodal models

Google is pushing hard into the local AI space, providing powerful tools that bypass cloud dependency and latency.

• –Native multimodality (including video, image, and native audio on edge models) combined with agentic workflow support makes these ideal for building autonomous, on-device agents.
• –The 26B Mixture of Experts (MoE) model is highly optimized, activating only 3.8 billion parameters to maximize speed without sacrificing workstation performance.
• –Delivering 128K context windows on the smaller E2B and E4B edge models is a massive leap for processing large documents on mobile devices and IoT hardware.
• –The permissive Apache 2.0 license signals a continuing commitment to the open-source developer community to counter closed-source alternatives.