Nemotron 3 Nano Omni

Nemotron 3 Nano Omni is positioned as a perception and context sub-agent inside enterprise agent systems, accepting text, images, video, and audio together and emitting text, so a single inference pass can reason across mixed modalities instead of stitching separate vision and speech pipelines. It is described as the first entry in the Nemotron multimodal series to natively support audio alongside the other input types, and the authors highlight real-world document understanding, long audio-video comprehension, and agentic computer use as the areas where it shows its strongest results compared to its predecessor, Nemotron Nano V2 VL.

The model is built on the Nemotron 3 Nano 30B-A3B backbone and pairs a hybrid MoE Transformer-Mamba design with Conv3D video layers and an Efficient Video Sampling (EVS) stage, using multimodal token-reduction techniques to cut inference latency and lift throughput relative to similarly sized models. Open weights are published on Hugging Face in BF16, with FP8 and FP4 variants also released alongside portions of the training data and code, making the system a practical fit for teams that want to self-host a compact multimodal reasoner or wire it into a larger agent stack where low-latency perception of documents, screens, and long audiovisual context matters.