NVIDIA just dropped something that could meaningfully change how AI agents are built – and it’s open for everyone to use. On April 28, 2026, the company unveiled Nemotron 3 Nano Omni, a multimodal model that does something most AI systems still struggle with: it sees, listens, reads, and thinks – all at once, all inside a single model.

To understand why that matters, it helps to think about how most AI agent systems work today. When a company builds an AI agent to handle customer support, for example, it typically strings together multiple models – one for understanding speech, one for analyzing screenshots, another for reading documents, and yet another to actually reason and respond. Every time data has to travel from one model to the next, the agent loses a little context, burns a little more time, and introduces another layer where things can go wrong. Multiply that across millions of interactions, and the cost – in compute, latency, and errors – gets enormous.

NVIDIA’s answer to this is Nemotron 3 Nano Omni, which brings vision, audio, and language understanding into one unified system. The model can take in text, images, audio, video, documents, charts, and even graphical user interfaces all at once, and respond with text – no handoffs, no separate perception pipelines. According to NVIDIA, this architectural consolidation enables AI systems to achieve up to 9x higher throughput compared to other open omni models with the same interactivity, which is a dramatic efficiency jump that directly translates to lower costs and better scalability for the businesses deploying these systems.

The technical architecture behind this is genuinely interesting. Nemotron 3 Nano Omni runs on a 30B-A3B hybrid Mixture-of-Experts (MoE) design – which means it has 30 billion total parameters, but only activates about 3 billion of them per token. This is the “nano” part of the name, and it’s what makes the model surprisingly lean despite its capabilities. It can run on roughly 25GB of RAM, which puts it within reach of high-end workstations, not just massive data center clusters. The hybrid MoE core cleverly combines Mamba layers – known for sequence and memory efficiency – with transformer layers for precise reasoning, delivering up to 4x improved memory and compute efficiency over a standard transformer approach.

Handling video is one of the trickier challenges for any multimodal model, and NVIDIA built specific machinery to address it. The model uses 3D convolutions to capture motion between video frames, combined with an inference-time Efficient Video Sampling (EVS) layer that compresses the dense stream of visual tokens from multiple frames into a compact set the language model can actually work with – without flooding its context window. For images, the system uses a dynamic resolution processing approach, supporting anywhere from 1,024 to 13,312 visual patches per image, which is the equivalent of handling images from 512×512 all the way up to roughly 1840×1840 pixels. And for the audio side, the model processes sound natively, allowing it to tie what was said in a conversation to what was shown on screen – without reducing either to a disconnected summary.

The model tops six multimodal benchmarks, including MMlongbench-Doc and OCRBenchV2 for complex document intelligence, and WorldSense, DailyOmni, and VoiceBench for video and audio understanding. On the MediaPerf benchmark – which tests how efficiently a model can process video at scale – Nemotron 3 Nano Omni achieved the highest throughput across every tested task and the lowest inference cost for video-level tagging, processing approximately 9.91 hours of video per hour versus around 3.8 hours per hour for Qwen3-VL, one of its closest competitors.

One of the most compelling real-world demonstrations of what this model can do comes from H Company, a startup building computer-use AI agents – software that can directly navigate and operate computers like a human would. Gautier Cloix, CEO of H Company, described the impact directly: “To build useful agents, you can’t wait seconds for a model to interpret a screen. 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.” H Company’s computer-use agent runs at a native input resolution of 1920×1080 pixels and showed strong preliminary results on the OSWorld benchmark, a standard test for agents navigating complex graphical interfaces.

The enterprise interest is already significant. Companies including Palantir, Foxconn, Docusign, Oracle, Infosys, and Dell Technologies are either already using or actively evaluating the model. Startups like Pyler are using it for video safety and content moderation at scale, while Eka Care is applying it to multimodal healthcare workflows in India. Applied Scientific Intelligence is using it as part of a scientific literature research agent. These aren’t frivolous experiments – they’re production-path deployments in regulated, high-stakes industries, which says something real about the confidence enterprises are placing in the model’s reliability.

What makes this particularly notable for developers and companies with strict compliance requirements is the openness of the release. NVIDIA published not just the model weights, but also the datasets and training recipes – giving organizations full transparency into what the model was trained on and how it behaves. That level of openness matters enormously for industries like healthcare, finance, and government, where AI systems often need to pass internal audits before deployment. Because the weights are fully open, companies can also customize the model using NVIDIA’s NeMo toolkit for their specific domain, fine-tuning it on proprietary data without shipping that data anywhere.

Deployment flexibility is another genuine selling point here. The model’s lightweight architecture means it can run on local systems like NVIDIA Jetson hardware for edge deployments, on NVIDIA DGX Spark and DGX Station for workstation-level use, and all the way up to data center and cloud environments. It supports hardware-optimized inference across NVIDIA’s Ampere, Hopper, and Blackwell GPU families, plays well with popular inference engines like vLLM and TensorRT-LLM, and supports FP8 and NVFP4 quantization for even more efficient runs on compatible hardware. As of launch, it’s available through Hugging Face, OpenRouter, build.nvidia.com, and more than 25 partner platforms.

In the broader context of where AI is heading, Nemotron 3 Nano Omni represents a meaningful design philosophy shift. Rather than throwing more parameters at the problem or scaling up context windows indefinitely, NVIDIA focused on consolidation, efficiency, and architectural cleverness. The Nemotron 3 family as a whole – covering Nano for perception, Super for high-frequency execution, and Ultra for complex long-horizon planning – has now passed 50 million downloads in the past year, suggesting developers are actively building with it rather than just experimenting. The Omni addition to the Nano tier extends that momentum into the multimodal and agentic domain at exactly the moment when the market needs it most – as the race to build AI systems that can genuinely perceive and act on the real world is heating up across every major industry.