Nvidia has released an open AI model that watches video, listens to audio and reads documents without calling on a second model, and it says folding those three jobs into one is worth up to nine times the throughput. Nvidia calls it Nemotron 3 Nano Omni. It carries 30 billion parameters, the weights are free to download, and it went up on Hugging Face, OpenRouter, the company's own build.nvidia.com and more than 25 partner platforms on April 28th, 2026.
The problem it aims at is plumbing. An agent asked to watch a screen recording, listen to a support call and read a log file typically runs a different model for each job and shuttles results between them. Nvidia's argument is that every handoff costs another round of inference, and that context leaks away at each transfer.
Omni is the shorthand for a model that accepts more than one kind of input. Here that means text, images, audio, video, documents, charts and graphical interfaces going in, with text coming out. Underneath sits a mixture-of-experts design, a way of building a very large model so that only part of it has to run for any given input. Picture a hospital that keeps an enormous roster of specialists on the payroll and a triage desk that wakes only the two who can treat the patient in front of them. Nvidia labels this particular arrangement 30B-A3B.
About that nine times. The figure is Nvidia's own, it compares total system throughput against other open omni models held at the same level of per-user interactivity, and the announcement page does not name the models it beat. It also claims the model tops six leaderboards for document intelligence and video and audio understanding without naming any of the six. An independent run on named benchmarks against named rivals would settle it; until then this is a vendor number carrying a vendor's methodology.
One named adopter is H Company, which builds agents that drive computers. “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.” Nvidia adds that H Company's agent feeds the model images at a native 1920 by 1080 resolution, and that preliminary evaluations on the OSWorld benchmark showed a significant leap in navigating complex interfaces. No number is attached to the leap.
Adoption, as Nvidia describes it, comes in two tiers. Aible, Applied Scientific Intelligence, Eka Care, Foxconn, H Company, Palantir and Pyler are listed as already adopting the model. Dell Technologies, Docusign, Infosys, K-Dense, Lila, Oracle and Zefr are listed as evaluating it, which is a much softer word.
The commercial logic is not subtle. Nvidia is handing out weights, datasets and training techniques at no charge, and every deployment target it names runs on its own silicon, from Jetson boards and DGX Spark desktops up to data center racks. The company says the wider Nemotron 3 family has passed 50 million downloads in the past year, a count only Nvidia is in a position to verify.
Whether one model can genuinely replace three in production gets settled in somebody's inference bill rather than in a launch post, and the enterprises listed as evaluating rather than adopting appear to know it.
Leave a Reply