NVIDIA Nemotron 3 Nano Omni (30B-A3B)

2026-04-28 · Publisher : NVIDIA · Domain : Artificial Intelligence

NVIDIA launches Nemotron 3 Nano Omni, an open model with 30 billion parameters of which about 3 billion are active, unifying vision, audio and language for agents, with a claimed 9 times higher throughput.

### Context Agents that use a computer or read documents often chain several separate models for vision, audio and text. NVIDIA offers a single model designed to serve as a perception sub-agent within larger systems, with its own license, distinct from that of the text models. ### What the model brings - Inputs: text, images, audio, video, documents, charts and graphical interfaces; text output with reasoning, JSON and tool calls. - A claimed throughput up to 9 times higher than other open omni models at equal interactivity. - Leading results on six document, video and audio understanding leaderboards, according to NVIDIA. - An improvement over Nemotron Nano VL V2: OCRBenchV2 67.04 (+18.26%), OSWorld 47.4 (+76.58%), chart reasoning 63.6 (+35.06%). - Reasoning enabled by default, which can be disabled with the enable_thinking parameter. ### Architecture and specifications Hybrid Mamba-2 Transformer MoE backbone of about 31 billion parameters (30B-A3B, about 3 billion active), CRADIO v4-H vision encoder and Parakeet audio encoder. Maximum context 256K tokens. Training: 354.6 million samples (about 717 billion tokens), drawn from public datasets, licensed data, NVIDIA in-house collections and synthetic annotations. License: NVIDIA Open Model Agreement (HF metadata), to be distinguished from the Nemotron Open Model License of the text models. Openness: weights, datasets and techniques announced as open; the extent of the published data was not detailed. ### Cost and access | Item | Detail | |---|---| | Official API | build.nvidia.com (NIM microservice), OpenRouter, more than 25 partner platforms; pricing not disclosed | | Weight size | 62 GB in BF16, 33 GB in FP8, 21 GB in NVFP4 (HF model card) | | Hardware | 1 x H100 80 GB minimum in BF16; H200, B200, RTX Pro 6000 recommended; Jetson Thor compatible | ### Benchmarks | Benchmark (published by NVIDIA) | Nano Omni | Gap vs Nemotron Nano VL V2 | |---|---|---| | OCRBenchV2 | 67.04 | +18.26% | | OSWorld (computer use) | 47.4 | +76.58% | | Chart reasoning | 63.6 | +35.06% | Source: Hugging Face model card nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 and NVIDIA Blog post. Comparison with NVIDIA's previous model, not with closed models. ### Versus closed models The sources reviewed do not compare Nano Omni to closed multimodal models. The only quantified gaps are against Nemotron Nano VL V2, an earlier open model. NVIDIA claims first place on six leaderboards without detailing the competitors in the excerpt reviewed, and a throughput 9 times higher than other open omni models. The proposition relies on self-hosting with a single H100 GPU for agent perception, whereas closed multimodal models are used through an API. The choice between open and closed therefore depends on the need for control over audio and video data. ### Key takeaways Nano Omni brings audio, video and document perception to the Nemotron family in the form of a single open model. Its own license should be checked before any commercial use.

Tags : LLM, Multimodal, Open Source, MoE, NVIDIA, Nemotron 3, NVIDIA Open Model Agreement