Key takeaways

• NVIDIA Nemotron 3 Nano Omni is an open multimodal model designed as a perception and context sub-agent for enterprise AI systems, unifying audio, images, video, and text processing.
• The model utilizes a Mixture of Experts (MoE) architecture, delivering up to 9x higher throughput than other open omni models and excelling at computer use, document intelligence, and audio-video reasoning tasks.
• The model shows strong multimodal accuracy, achieving 90.08% accuracy on the ARC-Challenge dataset for grade-school science questions, demonstrating strong comprehension and logical inference.
• A significant limitation is its severe struggle with the OSWorld (small) benchmark, achieving only 4.76% accuracy due to consistent failures in correctly parsing and executing GUI tool calls, indicating a major gap in real-world desktop interaction capabilities.
• The model also achieved 0% accuracy on the AIME 2025 benchmark, failing completely on advanced mathematics problems and highlighting limitations in complex, multi-step reasoning often found in such domains.

Technical specifications

Context length

256k tokens

Max output

—

Architecture

MoE

Modalities

text+image+audio+video->text

Parameters

~30B

Open weights

Yes

Licence

NVIDIA Open Model

Regions

USA

Vendor

Nvidia

Provider model key

nvidia/nemotron-3-nano-omni