What are the main differences between the models in the Gemma 4 family?

The Gemma 4 family is designed for different scales of deployment. Gemma-4-31B is a large, dense model for high-performance reasoning tasks in data centers. Gemma-4-26B-A4B is a Mixture-of-Experts (MoE) model also suited for demanding server environments. The Gemma-4-E4B and Gemma-4-E2B are smaller, more efficient models designed specifically for on-device and edge deployments like those on NVIDIA Jetson, offering full multimodal capabilities (text, audio, vision, video) in a compact form factor.

Bringing AI Closer to the Edge and On-Device with Gemma 4

NVIDIA has released the Gemma 4 family, a new suite of multilingual and multimodal AI models designed to operate across its entire hardware portfolio, from high-end Blackwell data center GPUs to consumer RTX systems and Jetson edge devices. The launch directly addresses growing industry demand for AI models that can be deployed locally, offering lower latency, improved cost efficiency, and enhanced security for on-premises requirements. The family includes four models with capabilities spanning complex reasoning, code generation, agentic tool use, and the ability to process interleaved text, image, audio, and video inputs.

The Gemma 4 series consists of four distinct models, including the company's first Mixture-of-Experts (MoE) architecture in the Gemma-4-26B-A4B. The flagship Gemma-4-31B is a dense model optimized for complex tasks, while the smaller Gemma-4-E4B and E2B variants are tailored for on-device and mobile applications with effective parameter counts of 4.5B and 2.3B respectively. For deployment, NVIDIA has coordinated support across popular frameworks like vLLM, Ollama, and llama.cpp, ensuring developers can run the models on hardware ranging from DGX Spark workstations to Jetson Orin Nano modules. Notably, the 31B model will have an NVFP4 quantized checkpoint available for Blackwell developers, enabling 4-bit precision to increase performance and lower token costs.

This release represents a strategic effort by NVIDIA to create a cohesive development ecosystem that bridges its diverse hardware offerings. By providing a single, scalable model family, the company enables developers to prototype on a local RTX machine, fine-tune on a DGX system using the NeMo framework, and deploy the same architecture to either enterprise servers with NVIDIA NIM microservices or to autonomous machines at the edge. The models are available under a commercial-friendly Apache 2.0 license, positioning them for broad adoption in both open-source projects and enterprise applications where on-device processing and data privacy are critical concerns.

By providing a single, scalable model family that runs consistently from consumer GPUs to data center servers and edge devices, NVIDIA is using Gemma 4 to unify its hardware ecosystem. This move incentivizes developers to remain within the NVIDIA stack for the entire AI lifecycle, from initial prototyping to final production deployment.