NVIDIA launches Nemotron3 Nano Omni model: capable of unified processing of video, audio, images, and text, enhancing multimodal reasoning efficiency

BlockBeats News, April 29 — NVIDIA officially launched Nemotron 3 Nano Omni, the newest member of the Nemotron 3 series, integrating unified multimodal reasoning into a single efficient open-source model. NVIDIA says that agentic systems typically need reasoning that performs a single perception-to-action loop across screens, documents, audio, video, and text, yet they still rely on fragmented chains of models—separate technical stacks for vision, audio, and text. This increases the number of reasoning hops and orchestration complexity, drives up inference costs, and at the same time weakens consistency of cross-modal context. Nemotron 3 Nano Omni is designed to replace this fragmented vision-language-audio technical stack, acting as a multimodal perception and contextual sub-agent (sub-agent) within agentic systems.

In terms of accuracy, Nemotron 3 Nano Omni achieves leading results on the document intelligence leaderboard, and it is also leading on the video and audio understanding leaderboards. On the open industry benchmark MediaPerf for video understanding models, Nemotron 3 Nano Omni achieves the highest throughput in each task and delivers the lowest inference cost in video-level annotation tasks.

In terms of performance, under a fixed per-user interaction threshold, for video inference, Nemotron 3 Nano Omni maintains higher overall system throughput and can achieve up to approximately 9.2 times the effective system capacity compared with other open-source omni models; for multi-document reasoning, it can achieve up to approximately 7.4 times the effective system capacity. NVIDIA says the model is intended to replace traditional multi-model concatenation architectures, reduce inference complexity and costs, and promote the application of multimodal AI in scenarios including finance, healthcare, scientific research, and media.