Google DeepMind's Gemma 4 Goes Multimodal and MoE

Google DeepMind has introduced Gemma 4, the latest generation of its open-weights model family, according to the Gemma 4 Technical Report. The release marks a notable architectural shift, moving the family toward a mixture-of-experts (MoE) design while broadening its native support for text, images, and reasoning-heavy tasks.

The headline changes are structural. Gemma 4 adopts an MoE approach, which activates only a subset of parameters per token to improve efficiency relative to dense models of comparable capacity. The family is also described as encoder-free for multimodal inputs, suggesting a more unified path for handling images and text without a separate vision encoder stage.

What's new in Gemma 4

• A mixture-of-experts architecture across the family
• Encoder-free multimodal handling for text and vision
• An optional thinking mode for step-by-step reasoning
• Support for long context

The thinking mode places Gemma 4 alongside a growing set of models that expose explicit reasoning behavior, letting developers trade extra compute for stronger performance on complex problems. Combined with long-context support, that positions the family for tasks that span large documents or multi-step workflows.

Gemma 4 continues to ship under the Gemma license, keeping the weights accessible for developers and researchers who want to build on or fine-tune the models directly. Why it matters: an open MoE multimodal family with reasoning controls narrows the gap between open weights and the capabilities typically reserved for closed frontier systems, giving teams more room to self-host and customize.