NVIDIA Nemotron 3 Nano Omni: 30B-A3B MoE Multimodal Model with 9× Higher Throughput for AI Agents

NVIDIA introduced Nemotron 3 Nano Omni on April 28, 2026 — an open-source multimodal model that unifies vision, audio, and language in a single architecture for AI agents. The model arrives as a 30B-A3B hybrid mixture-of-experts (MoE) with a 256K context window and throughput 9× higher than other open omni models at the same level of interactivity. Its positioning is a direct challenge to recent multimodal releases from Mistral, Meta, and Alibaba.

What Modalities Does Nemotron Nano Omni Process?

The model accepts and processes six input types in a single architecture: text, images, audio, video, documents with charts, and GUI screenshots. Specific technical components include Conv3D and an EVS (efficient video scaling) module for video, and the system has been tested on 1920×1080 resolution displays for GUI navigation. It is designed primarily for AI agents that need to combine UI observation, document reading, and user conversation in the same workflow.

What Does 9× Higher Throughput Mean for Inference?

NVIDIA claims that Nano Omni achieves 9× more tokens per second compared to other open omni models while maintaining identical interactivity. In practice, this means that agentic workflows previously bottlenecked by the latency of multimodal processing — for example, reading hundreds of pages of a document while simultaneously clicking through a GUI — can now run in real time. The model currently leads in six categories on public leaderboards for document, video, and audio understanding, although NVIDIA does not cite specific benchmark numbers in the announcement.

Where Is It Available and Who Is Already Using It?

The model is available via Hugging Face, OpenRouter, NVIDIA’s build.nvidia.com portal, and more than 25 partner platforms. Active early users include Aible, Applied Scientific Intelligence, Eka Care, Foxconn, H Company, Palantir, and Pyler. Dell Technologies, Docusign, Infosys, Oracle, and Zefr are in the evaluation phase. The broad distribution and list of corporate users suggest that NVIDIA prepared the model for immediate enterprise deployment, not just research testing.