NVIDIA releases Nemotron 3 Nano Omni, an open 30B multimodal model for AI agents with 9x throughput over comparable open omni models

NVIDIA released Nemotron 3 Nano Omni on April 28, 2026, an open multimodal model designed for AI agent deployments that unifies video, audio, image, and text processing in a single architecture. The model tops six leaderboards for complex document intelligence and video and audio understanding, while delivering 9x higher throughput than other open omni models with comparable interactivity.

What's new

Nemotron 3 Nano Omni uses a 30B-A3B hybrid mixture-of-experts (MoE) architecture with a 256K context window. The "30B-A3B" notation means the model has 30 billion total parameters but activates only 3 billion per forward pass, enabling high throughput at a fraction of the compute cost of dense models.

The model supports a wide input surface:

• Inputs: Text, images, audio, video, documents, charts, and graphical user interfaces
• Output: Text

According to NVIDIA, Nemotron 3 Nano Omni "brings these capabilities together into one system, enabling agents to deliver faster, smarter responses with advanced reasoning across video, audio, image and text."

The model ships with fully open weights, datasets, and training recipes, making it customizable for domain-specific deployments. It is available on Hugging Face, OpenRouter, build.nvidia.com as an NVIDIA NIM microservice, and across more than 25 partner platforms including NVIDIA Cloud Partners.

Early adopters announced at launch include Aible, Palantir, and Foxconn.

Context

Nemotron 3 Nano Omni is part of NVIDIA's Nemotron 3 model family, which targets developer and enterprise deployment of open, customizable AI. Earlier models in the family focused on language; the Nano Omni extends the family to full omni-modal support, including audio and video — categories where open models have historically lagged behind proprietary systems.

The 9x throughput claim over other open omni models positions Nemotron 3 Nano Omni as an infrastructure-efficiency argument rather than a pure benchmark argument. For teams running high-traffic agent pipelines, throughput is often the binding constraint, and a 9x gap is a significant operational difference.

NVIDIA's choice to release with fully open weights, training data, and recipes (not just model weights) also distinguishes it from models released under restrictive access terms. The Apache-style openness allows commercial use without licensing friction.

Why it matters

For agent builders, the multimodal bottleneck has been the absence of a single open model that handles all input types competently. Most production pipelines route text, image, audio, and video through different specialized models, adding latency and system complexity. A single open omni model with strong throughput characteristics reduces that complexity while keeping inference costs predictable.

The 256K context window also enables longer in-context reasoning over documents, transcripts, and video content — the kind of context AI agents need for enterprise workflows where inputs span multiple modalities and can be lengthy.