Breaking: NVIDIA unveils Nemotron 3 Nano Omni
April 28, 2026 — NVIDIA today released Nemotron 3 Nano Omni, an open multimodal model that unifies vision, audio, and language into a single system, enabling AI agents to process video, audio, images, and text up to 9 times more efficiently than existing solutions.
The model, available immediately on Hugging Face, OpenRouter, and build.nvidia.com, marks a leap in agentic AI performance: it tops six leaderboards for document intelligence and multimodal understanding while cutting inference costs by up to 90% compared to current open omni-models.
“You can’t wait seconds for a model to interpret a screen,” said Gautier Cloix, CEO of H Company, an early adopter. “By building on Nemotron 3 Nano Omni, our agents can rapidly interpret full HD screen recordings — something that wasn’t practical before. This isn’t just a speed boost: It’s a fundamental shift in how our agents perceive and interact with digital environments in real time.”
The Efficiency Problem in Multimodal Agents
Most AI agent systems today rely on separate models for vision, speech, and language, passing data from one to another. This fragmented approach introduces repeated inference passes, fragments context across modalities, and increases latency and cost over time.
Nemotron 3 Nano Omni consolidates these tasks into a single model — a 30 billion parameter, 3 billion active hybrid Mixture-of-Experts (MoE) architecture with Conv3D, Event-based Vision Sensors (EVS), and 256K context window. It accepts text, images, audio, video, documents, charts, and graphical interfaces as input, and outputs text.
Key Specifications
- What it is: An open, omni-modal reasoning model — the highest-efficiency open multimodal model of its kind with leading accuracy.
- What it handles: Text, images, audio, video, documents, charts and graphical interfaces (input); text (output).
- Who it’s for: Enterprises and developers building fast, reliable agentic systems that need a multimodal perception sub-agent.
- How it works: Functions as the “eyes and ears” in a system of agents, working alongside models like Nemotron 3 Super and Ultra or other proprietary models.
- Why it matters: Leading multimodal accuracy and 9x higher throughput than other open omni models with the same interactivity, resulting in lower cost and better scalability without sacrificing responsiveness.
- Architecture: 30B-A3B hybrid MoE with Conv3D, EVS, 256K context.
- Availability: April 28th, 2026 via Hugging Face, OpenRouter, build.nvidia.com and 25+ partner platforms.
Background
AI agents for customer support, finance, and other sectors traditionally juggle separate models for vision, speech, and language. Each model introduces latency and context fragmentation — for example, a customer support agent processing a screen recording while analyzing uploaded call audio and checking data logs would require multiple inference steps across separate systems.
Nemotron 3 Nano Omni eliminates this overhead by combining vision and audio encoders within one architecture. It achieves up to 9x higher throughput than competing open omni-models, making real-time multimodal interactions practical at scale.
What This Means
For enterprises, the model provides a production-ready path to building more accurate and faster AI agents without the cost and complexity of managing multiple models. Early adopters include Aible, Applied Scientific Intelligence (ASI), Eka Care, Foxconn, H Company, Palantir, and Pyler, with Dell Technologies, Docusign, Infosys, K-Dense, Lila, Oracle, and Zefr evaluating the model.
“This isn’t just a speed boost — it’s a fundamental shift,” Cloix added. “Our agents can now interpret full HD screen recordings in real time, something that was impractical before. The efficiency gains are transformative for real-time digital environments.”
The model is open and available under a permissive license, giving developers full deployment flexibility and control. With its leading accuracy and low cost, Nemotron 3 Nano Omni sets a new efficiency frontier for open multimodal models.