NVIDIA Releases Nemotron 3 Nano Omni: Open 30B Multimodal Model
NVIDIA has open-released Nemotron 3 Nano Omni — a 30B MoE / 3B active parameter model designed as a single unified architecture handling video, audio, image, and text input with reasoning output. NVIDIA reports a 9× system capacity improvement for video reasoning versus predecessor models. The release includes full model weights, datasets, and training recipes, making it one of the most complete open multimodal releases of 2026. The full version is 66.1 GB.
Why It Matters
A fully open multimodal model with training recipes gives practitioners a replicable foundation for video-reasoning applications without proprietary model lock-in — directly challenging closed leaders in the video-AI segment.