Nemotron 3 Nano Omni

NVIDIA's 31B mixture of experts (3B active) for video, audio, images and text, with a 256K context and commercial use allowed.

In short. Nemotron 3 Nano Omni can watch a video with its soundtrack, listen to audio, look at images and read text, and it answers in text. It only activates about 3B of its 31B parameters per token, so it is cheaper to run than its size suggests, and it has a thinking mode that is on by default.

Summary

Nemotron 3 Nano Omni is NVIDIA's multimodal reasoning model that takes video, audio, images and text and answers in text. It combines the Nemotron 3 Nano backbone, a 31B Mamba2-Transformer hybrid mixture of experts with about 3B active parameters per token, with the C-RADIOv4-H vision encoder and a Parakeet speech encoder, which brings the checkpoint to 33.0B parameters. The card gives a 256K-token context, English-only language support, up to 2 minutes of mp4 video and up to 1 hour of wav or mp3 audio, and reasoning that is on by default and switched with enable_thinking. NVIDIA aims it at video and speech analysis, document intelligence, GUI and agent workflows and transcription, and it supports tool calling, JSON output and word-level timestamps.

NVIDIA's own table compares it with Nemotron Nano VL V2: 67.04 against 54.8 on OCRBench v2 (English), 47.4 against 11.1 on OSWorld, 63.6 against 41.3 on Charxiv reasoning, 57.5 against 38 on MMLongBench-Doc, 83.95 against 78.3 on CVBench2D and 82.8 against 75.5 on MathVista-mini. It lists Video-MME at 72.2, WorldSense at 55.4, Daily Omni at 74.52 and voice interaction at 89.39 without a comparison. The FP8 and NVFP4 builds stay within about 0.4 points of BF16 on the mean of nine non-speech benchmarks in non-reasoning mode, by NVIDIA's measurement. NVIDIA says the model was improved using Qwen3-VL-30B-A3B-Instruct, Qwen3.5-122B-A10B, Qwen3.5-397B-A17B, Qwen2.5-VL-72B-Instruct and gpt-oss-120b.

The weights come as BF16 (66 GB), FP8 (35 GB) and NVFP4 (22 GB), and NVIDIA lists the minimum GPU as one 80 GB H100 for BF16, one 48 GB L40S for FP8 and one 32 GB RTX 5090 for NVFP4. The card requires vLLM 0.20.0, with an audio extra if any audio is used, and also lists SGLang for BF16, TensorRT-LLM, llama.cpp, Ollama and Jetson Thor. The target system is Linux, and the API does not accept raw PDFs, so pages must be rendered to PNG first. The license is the NVIDIA Open Model Agreement, which allows commercial use. In the 30 days to 11 November 2026, Hugging Face counted about 116,000 downloads for the BF16 repository, 91,000 for FP8 and 50,000 for NVFP4.

Good for

  • Analyzing meetings, lectures and training videos where the picture and the speech have to be understood together
  • Document, chart and screenshot question answering and OCR under a license that allows commercial use
  • Agents that operate a computer screen, since NVIDIA reports 47.4 on OSWorld
  • One-GPU deployments, down to an RTX 5090 with the NVFP4 build

Watch out for

  • It supports English only, video up to 2 minutes and audio up to 1 hour, and it answers in text, not speech or images
  • Reasoning is on by default, so chain-of-thought traces appear in the output unless you turn thinking off, and vLLM samples only about 32 video frames unless you set the frame options at launch
  • The benchmark table is NVIDIA's own and compares against an older NVIDIA model, so check how it does against Gemma 4 or Qwen on your data
  • It needs vLLM 0.20.0 or a listed alternative, custom code and Linux, and PDFs have to be rendered into page images yourself

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Nano Omni 30B-A3B (NVFP4) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Nano Omni 30B-A3B (BF16) BF16 31B 3B active 71 GB 36 GB 80 GB Workstation
Nano Omni 30B-A3B (FP8) FP8 31B 3B active 71 GB 36 GB 48 GB Workstation
Nano Omni 30B-A3B (NVFP4) NVFP4 31B 3B active 71 GB 36 GB 32 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 32 GB of GPU memory.

VRAM calculator

The license in plain words

A permissive, Apache-style agreement from NVIDIA, released on 2 April 2026. NVIDIA confirms that its works are commercially usable, that you may create and distribute derivative works, and that it claims no ownership of outputs. The main duties are to pass the agreement on and keep notices.

More about NVIDIA Open Model

See also

Frequently asked questions

What is Nemotron 3 Nano Omni?
Nemotron 3 Nano Omni can watch a video with its soundtrack, listen to audio, look at images and read text, and it answers in text. It only activates about 3B of its 31B parameters per token, so it is cheaper to run than its size suggests, and it has a thinking mode that is on by default.
Can I use Nemotron 3 Nano Omni commercially?
Yes. Nemotron 3 Nano Omni is released under NVIDIA Open Model, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Nemotron 3 Nano Omni open source?
Nemotron 3 Nano Omni is open-weight, not open source in the strict sense. You can download the model files, but the NVIDIA Open Model adds conditions that the Open Source Initiative definition does not allow.
Can I run Nemotron 3 Nano Omni on my own computer?
Yes, on a good graphics card. The smallest variant, Nano Omni 30B-A3B (NVFP4), needs about 32 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
When was Nemotron 3 Nano Omni released?
Nemotron 3 Nano Omni was released by NVIDIA on April 28, 2026.
How long a context window does Nemotron 3 Nano Omni have?
Nemotron 3 Nano Omni supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Nemotron 3 NVIDIA 31.6B – 120B 12B active 1M NVIDIA Nemotron (Permissive) 21 GB Gaming GPU
Gemma 4 Google DeepMind 5.1B – 30.7B 256K Apache 2.0 (Permissive) 3.52 GB Laptop or phone
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Mage-VL Microsoft 4.74B 256K Apache 2.0 (Permissive) 3.27 GB Laptop or phone