Nvidia Releases Nemotron 3 Nano Omni Multimodal Model

Nvidia has released Nemotron 3 Nano Omni, an open-source multimodal AI model that integrates vision, audio, language, and other perception capabilities into a single architecture to power efficient AI agent systems.

The 30 billion-parameter model activates only about three billion parameters per inference through a hybrid mixture-of-experts design, combining Mamba-Transformer layers, a CRADIOv4H vision encoder, and a ParakeetTDT audio encoder. It processes inputs including text, images, audio, video, documents, charts, and graphical user interfaces, producing text outputs. Supporting a context window of up to 256,000 tokens and primarily English language processing, the model targets enterprise developers building agentic workflows such as document intelligence, computer-use agents, audio-video reasoning, and voice interactions.

Nvidia states that Nemotron 3 Nano Omni delivers nine times higher throughput than comparable open multimodal models at the same interactivity level, along with leading performance on benchmarks including OCRBenchV2, MMLongBenchDoc, WorldSense, VoiceBench, and OSWorld. On OSWorld, a GUI agent benchmark, it achieved 47.4 percent accuracy, a significant...