Nvidia Unveils Nemotron 3 Nano Omni To Unite Vision, Audio And Language

Nvidia has launched Nemotron 3 Nano Omni, an open "omni‑modal" reasoning model designed to process video, audio, images and text within one architecture, reducing the hand‑offs between separate vision, speech and language models that add latency and lose context.

The company says the model sets a new efficiency mark for open multimodal systems, combining strong perception accuracy with lower cost, and topping six leaderboards spanning complex document intelligence plus video and audio understanding. According to Nvidia, systems built on the model can achieve up to ninefold higher throughput than other open omni models with similar interactivity.

Nemotron 3 Nano Omni integrates vision and audio encoders inside a 30B‑A3B hybrid mixture‑of‑experts design, eliminating the need for standalone perception components and improving inference efficiency at scale. Its architecture includes Conv3D, EVS and a 256K context window, and it accepts inputs across text, images, audio, video, documents, charts and graphical interfaces, producing text outputs.

In multi‑agent setups, the model can act as the "eyes and ears", working alongside proprietary cloud models or other open Nemotron models — such as Nemotron 3 Super for high‑frequency execution and Nemotron 3 Ultra for complex planning — to power sub‑agents for computer use, document intelligence and audio‑video reasoning.

Early use cases include computer‑use agents that navigate GUIs and reason over on‑screen content at native 1920×1080 resolution, where preliminary OSWorld benchmark results indicate gains in handling complex interfaces; document intelligence that coherently interprets charts, tables, screenshots and mixed media for enterprise analysis and compliance; and audio/video understanding that maintains context across what was shown and said in a single reasoning stream.

"To build useful agents, you can't wait seconds for a model to interpret a screen," said Gautier Cloix, CEO of H Company. "By building on Nemotron 3 Nano Omni, our agents can rapidly interpret full HD screen recordings — something that wasn't practical before. This isn't just a speed boost: It's a fundamental shift in how our agents perceive and interact with digital environments in real time."

Adopters already include Aible, Applied Scientific Intelligence (ASI), Eka Care, Foxconn, H Company, Palantir and Pyler, with Dell Technologies, Docusign, Infosys, K‑Dense, Lila, Oracle and Zefr evaluating the model.

Nvidia is releasing Nemotron 3 Nano Omni with open weights, datasets and training methods, enabling organisations to customise, evaluate and optimise the model for domain‑specific tasks using tools such as Nvidia NeMo. Because the Nemotron family is open, deployments can be aligned with regulatory, sovereignty and data localisation requirements.

Availability starts 28 April 2026 via Hugging Face, OpenRouter and build.nvidia.com as an Nvidia NIM microservice, and through a wide ecosystem of Nvidia Cloud Partners, inference platforms and cloud providers. The lightweight, open design supports consistent deployment from local systems like Nvidia Jetson hardware, Nvidia DGX Spark and DGX Station to data centre and cloud environments.

Nvidia says the Nemotron 3 family — spanning Nano, Super and Ultra — has been downloaded more than 50 million times in the past year, with Omni extending the line into multimodal and agentic workloads.

www.nvidia.com/en-gb/