Google Unveils Gemma 4 Model Lineup with Dense, MoE, and Multimodal Variants

Google has disclosed the detailed composition of its Gemma 4 model family through its AI developer documentation page. The announcement includes three main architectural variants: dense, mixture-of-experts (MoE), and unified multimodal models.

Architectural Variants

Dense models follow the traditional transformer structure, with all parameters activated during inference. This provides predictable latency and consistent throughput.

MoE architectures activate only a subset of expert subnetworks depending on the input, reducing the number of active parameters relative to the total parameter count. The routing mechanism selects expert combinations based on input tokens.

The unified multimodal model is designed to process text and images within a single architecture. It can support tasks such as visual question answering, document understanding, and multimodal retrieval.

Developer Ecosystem

The Gemma series has drawn attention in the open-weight model market, and the fourth-generation lineup expands the available options. Dense models are highly compatible with standard inference frameworks and are easier to integrate into existing pipelines.

MoE models require runtimes that support routing logic and expert load balancing. Multimodal variants place greater emphasis on input pipeline design, including image preprocessing, resolution adjustment, and text-image alignment.

Competitive Landscape

The open-weight model market includes Meta's Llama series, Mistral AI's model family, and Alibaba's Qwen lineup. Gemma 4's MoE variant may be compared with other MoE models, while the multimodal model may be evaluated alongside other multimodal offerings.

Licensing and Deployment

Gemma models are generally distributed under licenses that permit commercial use, but specific terms should be checked in the model cards and terms of service. MoE and multimodal variants may have higher inference memory requirements.

Google's official documentation is expected to include recommended hardware specifications, batch size settings, and inference optimization guides for each variant. The currently disclosed information confirms the existence of the model variants but does not specify parameter counts, benchmark performance, training data composition, or release schedules.