Overview
Run NVIDIA Nemotron 3 Nano Omni 30B A3B Reasoning with high-performance inference. Deploy NVIDIA’s multimodal reasoning model for text, image, video, and audio input in a few clicks.
Explore additional models
Frequently
asked questions
Nemotron 3 Nano Omni 30B A3B Reasoning is an open multimodal model from NVIDIA for reasoning across text, images, video, and audio. It was built on the Nemotron 3 Nano 30B A3B backbone with a C-RADIOv4-H vision encoder and a Parakeet-TDT audio encoder. It is the first Nemotron model with audio understanding, and it suits video and image analysis, transcription and summarization, document Q&A over scans, and agents that interpret what they see or hear. It is available for Serverless Inference in Crusoe Intelligence Foundry.
Nemotron 3 Nano Omni 30B A3B Reasoning adds image, video, and audio input to the text-only Nemotron 3 Nano 30B A3B, while sharing the same roughly 3B-active hybrid Mamba-Transformer backbone and 262,144-token context. Omni carries extra encoders and processes media tokens, so it is slower and costs more per token than Nano. Use Nano for text-only work and Omni when inputs include media.
Nemotron 3 Nano Omni 30B A3B Reasoning on Serverless Inference is billed per token, with separate rates for uncached, cached, and output tokens, and no minimum commitment. Media inputs are converted to tokens, so longer videos and higher-resolution images use more input tokens. Access it via an OpenAI-compatible API in Crusoe Intelligence Foundry. Current rates are shown in the pricing module above.
Nemotron 3 Nano Omni reasons across mixed media in a single pass, working step by step over images, video frames, and audio rather than handling each separately. Its 262,144-token window fits long transcripts and multi-frame video in one request, which suits video analysis, document Q&A over scans, and agents that act on what they observe.
Nemotron 3 Nano Omni is available for Serverless Inference with no infrastructure setup required. It is not currently offered for Self-Serve Deployments or Serverless Fine-Tuning. For an SLA-backed dedicated endpoint, contact us about Tailored Deployments.
The “A3B” in the model name stands for about 3 billion active parameters. The language backbone is a mixture-of-experts with roughly 30B total parameters, but it routes each token through only about 3B of them. The model holds 30B-class knowledge in memory while computing like a much smaller one, so multimodal reasoning stays fast and cost-efficient at high request volumes.


%204.png)
