With Nemotron 3 Nano Omni, Nvidia reveals what really goes into a modern multimodal model

Nvidia launched Nemotron 3 Nano Omni, a 30B-parameter open multimodal model that handles text, images, video, and audio in one system.

The model is aimed at agentic use cases and shows strong benchmark results across multimodal and OCR-related tasks.

Nvidia also disclosed unusual detail on training sources, synthetic data, partial datasets, and RL tooling, including outputs from other models.

The release adds transparency around how large multimodal models are built and could shape future model development and auditing.