Nemotron Nano Omni
NVIDIANemotron Nano Omni is NVIDIA's omni-modal reasoning model, a 30B-total, 3B-active mixture-of-experts that understands video, audio, images and text and answers in text. It reads video with its soundtrack, transcribes speech with word-level timestamps, parses long documents and drives computer interfaces, with tool calling in a 256K context. NVIDIA ships BF16, FP8 and NVFP4 builds.
Openness
3 high confidence- weights
- open(safetensors on the Hub, ungated)
- data
- partial(Nemotron-Image-Training-v3 is released under CC-BY-4.0
- code
- closed(serving and deployment instructions only)
- license
- NVIDIA-Open-Model-Agreement("Works are commercially usable"
The weights are published under NVIDIA's Open Model Agreement, which permits commercial use and derivatives. NVIDIA releases one of its image training sets, but the full training corpus includes licensed and internal data and no training code is published. NVIDIA-Open-Model-Agreement restricts neither who may use it nor at what scale, so the license tier is permissive_non_osi.
- https://huggingface.co/api/datasets/nvidia/Nemotron-Image-Training-v3 recorded 2026-09-27
Nemotron-Image-Training-v3: public, ungated, cc-by-4.0, image-centric multimodal training data
- https://huggingface.co/api/models/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16?expand[]=downloads&expand[]=cardData&expand[]=gated&expand[]=createdAt&expand[]=lastModified&expand[]=siblings recorded 2026-09-27
Hub metadata: gated false, safetensors shards, license_name nvidia-open-model-agreement, datasets nvidia/Nemotron-Image-Training-v3
- https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/main/README.md recorded 2026-09-27
Training data: 354,587,705 items across 1,395 datasets including "licensed third-party data, NVIDIA-internal collections"; the card gives vLLM, TensorRT-LLM and SGLang serving steps and no training scripts
- https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/ recorded 2026-09-27
"NVIDIA Open Model Agreement", release date April 2, 2026: "Works are commercially usable. You are free to create and distribute Derivative Works."
Adoption
4 high confidenceHugging Face downloads over the trailing 30 days, summed across NVIDIA's BF16, FP8 and NVFP4 builds of the model. Most of it is the quantized builds rather than the BF16 original.
- https://huggingface.co/api/models?author=nvidia&search=Omni&limit=100 recorded 2026-09-27
The three Nemotron-3-Nano-Omni-30B-A3B-Reasoning builds: BF16 351,706, NVFP4 647,309 and FP8 1,073,945 downloads in the trailing 30 days, 2,072,960 in all
Capability
5 high confidenceNemotron Nano Omni understands audio and video natively together with images and text, and also operates computer interfaces, scoring 47.4 on OSWorld. It sits level with Qwen-Omni, with one difference: it answers in text only, where Qwen-Omni also speaks.
- https://huggingface.co/nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16/raw/main/README.md recorded 2026-09-27
"Modalities (in): Video, Audio, Image, Text" / "Modality (out): Text"; "Supports tool calling"; OSWorld 47.4, Video MME 72.2, OCRBenchV2 (EN) 67.04
Verified 2026-09-27