NVIDIA AI Unveils Nemotron 3 Nano Omni, a Multimodal MoE for Agentic Workloads

NVIDIA AI has unveiled Nemotron 3 Nano Omni, a cutting-edge open-source multimodal Mixture-of-Experts (MoE) model designed for agentic workloads. With 30 billion total parameters but only 3 billion active per token, it achieves high efficiency for real-time tasks. The model supports a 256K context window and processes text, image, video, audio, and documents, making it versatile for complex AI agents that interact across multiple modalities.

This release, dated April 29, 2026, strengthens NVIDIA's position in open-source AI, offering developers a powerful tool for building autonomous agents. The MoE architecture allows selective activation of parameters, reducing computational costs while maintaining performance. Nemotron 3 Nano Omni is optimized for tasks like multimodal reasoning, document analysis, and real-time video understanding, enabling applications from customer service bots to advanced research assistants. Its 256K context window handles long-form content like entire documents or video streams, crucial for agentic workflows that require sustained attention. By open-sourcing the model, NVIDIA aims to accelerate innovation in AI agents, providing a robust foundation for enterprises and researchers to customize and deploy.