NVIDIA unveils Nemotron 3 Nano Omni for long documents, audio, video and AI agents

On April 28, 2026, NVIDIA introduced Nemotron 3 Nano Omni, a long-context multimodal model for documents, audio, video and agentic scenarios in interfaces. The company is betting on practical tasks: from parsing complex PDFs and screen recordings to speech recognition and reasoning across several data types at once.

What tasks does it cover

Nemotron 3 Nano Omni is positioned not as just an OCR model or yet another VLM for images. NVIDIA describes it as a general-purpose system for five classes of workloads: real-world document analysis, automatic speech recognition, understanding of long videos with audio, assistance in GUI scenarios, and general multimodal reasoning.

This is not about short demo examples, but about documents with tables, formulas, cross-page references, slides, screenshots and screen recordings with voice commentary. In document scenarios, the model, according to the company, handles files of more than 100 pages and is expected to retain both fine details and the overall structure at the same time. For audio and video, the emphasis is on long material: training videos, meetings with slides, product demos and support recordings. For agentic tasks, working with screenshots and interface state matters: the model can interpret what it sees on screen and help choose the next action.

• Multi-page contracts, reports and technical documents
• Screen recordings and tutorials with voice narration
• Recognition of long speech with noise, accents and multiple speakers
• Analysis of GUIs and screenshots for computer-use scenarios

What's inside the model

The architecture is built around the Nemotron 3 Nano 30B-A3B language backbone and two specialized encoders: C-RADIOv4-H for visual data and Parakeet-TDT-0.6B for audio. The link between modalities and the LLM is implemented through lightweight projectors that merge everything into a single token sequence.

Inside the backbone, NVIDIA uses a hybrid approach: 23 Mamba layers for long context, 23 MoE layers with 128 experts and top-6 routing, and 6 attention layers for global connections and complex reasoning.

There is a particular focus on efficiency with dense visual data. Instead of the tiling used in the previous version, the model gets dynamic resolution at the native aspect ratio: from 1024 to 13312 visual patches are allocated per image. For video, two compression mechanisms are used. Conv3D merges pairs of adjacent frames before they reach the ViT, while EVS at inference time discards static tokens and keeps only dynamic changes.

For audio, the move to native input matters: the model works not only with a transcript but with the audio track itself. It was trained on segments of up to 20 minutes, and the LLM's overall context is stated to exceed five hours.

Results and availability

In the published benchmarks, Nemotron 3 Nano Omni improves markedly over Nemotron Nano V2 VL and often outperforms Qwen3-Omni 30B-A3B. According to NVIDIA, the model scores 57.5 on MMLongBench-Doc versus 38.0 for the previous version, 65.8 on OCRBenchV2-En and 63.6 on CharXiv reasoning. On GUI tasks it scores 47.4 on OSWorld versus 11.0 for the previous model, and in multimodal video it reaches 72.2 on Video-MME, 55.4 on WorldSense and 74.1 on DailyOmni. For audio, the company reports 89.4 on VoiceBench and a 5.95 WER on HF Open ASR, where lower is better.

Cost and speed are an equally important point for developers. NVIDIA cites a 7.4x gain in system efficiency in multi-document scenarios and 9.2x in video use cases compared with other open multimodal models at comparable interactivity. The company also claims up to 2.9x higher single-stream reasoning speed in multimodal tasks.

Checkpoints are already available on Hugging Face in BF16, FP8 and NVFP4 formats, so the model can be tried not only as a research release but also as a base for applied pipelines.

What this means

NVIDIA is clearly heading not toward another showcase demo but toward working enterprise scenarios, where a system must read long documents, understand voice, see the interface and hold a large context at once, without a sharp rise in cost. If the claimed metrics hold up in real integrations, Nemotron 3 Nano Omni could become a strong open-weight candidate for document AI, video understanding and computer-use agents.