NVIDIA introduces the Nemotron 3 Nano Omni model, offering a 9x improvement in multimodal inference efficiency.

NVIDIA announced the release of the open-source multimodal model Nemotron 3 Nano Omni on its X platform. This model employs a 30B-A3B hybrid expert architecture (MoE), supports 256K context, and can uniformly handle video, audio, image, and text input. Compared to similarly interactive open-source all-around models, this model offers up to 9x higher throughput, significantly reduces inference costs, and improves scalability. Nemotron 3 Nano Omni is currently available on Hugging Face, OpenRouter, and NVIDIA NIM, and has been adopted by companies such as Aible, Applied Scientific Intelligence, and H Company.