Nvidia releases Nemotron 3 Nano Omni, an open multimodal model for edge agents

On April 28, 2026, Nvidia unveiled Nemotron 3 Nano Omni, an open multimodal model built for autonomous AI agents on edge devices. It is not just another release for the CUDA ecosystem: the company is signaling that it wants to earn money not only from hardware, but from the models themselves.

What the model can do

Nemotron 3 Nano Omni combines understanding of text, images, audio and video in a single architecture. The model also works with documents, charts and graphical interfaces, and outputs a text response. In effect, Nvidia is offering not a bundle of separate models for vision, speech and documents, but one shared engine for tasks where an agent must simultaneously see a screen, read a file, hear a voice command and react without extra delays between services.

The key idea is that the model is large in total size but relatively light in operation. Nemotron 3 Nano Omni has 30 billion parameters, but only 3 billion are activated at each inference step thanks to its mixture-of-experts architecture. Nvidia claims this approach delivers up to nine times higher throughput than comparable open multimodal models, and that the model leads in six benchmarks for document, video and audio tasks. The base text component was trained on 25 trillion tokens and supports a context window of up to 256K tokens.

Internally, the model uses a hybrid Mamba-Transformer design. According to Nvidia, it combines 23 Mamba-2 layers, 23 mixture-of-experts layers and six grouped-query attention layers. Each token is routed to only six of 128 experts plus a shared expert, so compute does not balloon. For video, the model uses three-dimensional convolutions that capture motion between frames rather than treating a clip as a set of static images. This engineering is what is supposed to make the model suitable for real-time agents on a single GPU.

• 30 billion parameters in total, 3 billion active at inference
• can run on a single GPU, without a server cluster
• commercial use is permitted under the Nvidia Open Model Agreement
• the model is available on Hugging Face and through Nvidia NIM
• inputs include text, images, audio, video, documents and graphical interfaces

Why Nvidia is doing this

Over the past two years, Nvidia has won primarily as an infrastructure supplier: GPUs, networking, CUDA and all the software around them. But the Nemotron family has already become a separate line of business, and now the company is making a bolder move, promoting its own model as a ready-made foundation for industrial AI agents.

The logic is simple: if the model is optimized for Nvidia hardware and the hardware for the Nvidia model, the company controls almost the entire stack, like Google, Amazon or Microsoft in their cloud ecosystems. That is why the release is presented not as a demonstration of lab capabilities, but as a product for deployment.

Nvidia names Foxconn, Palantir, Aible, ASI, Eka Care and H Company among early users and partners; Dell, DocuSign, Infosys, Oracle and Zefr are evaluating the model for production. The use cases are also not consumer-oriented: visual inspection in factories, document processing, voice agents and screen understanding for computer-use systems. The model can be deployed through Amazon SageMaker JumpStart, OpenRouter, vLLM, SGLang, Ollama, llama.cpp and TensorRT-LLM, meaning Nvidia wants it available in any familiar stack.

Against its competitors, the bet looks fairly well aimed. Google has the multimodal Gemini and the on-device Gemini Nano, Meta has the strong Llama lineup, and OpenAI remains the commercial benchmark with its GPT models. But Nvidia is trying to assemble a rare combination of four properties at once: multimodality, open weights, a commercial license and operation on edge hardware without cloud scale. If it works, the company can capture value at three levels at once: hardware, inference tooling and the model itself.

What this means

Nemotron 3 Nano Omni is a bet that the next wave of AI agents will run not only in the cloud but locally, closer to data, cameras, microphones and corporate interfaces. If Nvidia confirms the stated figures in real deployments, it will become not just a supplier of "shovels" for the AI boom, but one of the strongest players in the market for the models themselves.

*Meta has been designated an extremist organization and is banned in Russia.