nemotron-3-nano-30b-a3b

Nemotron 3 Nano 30B-A3B is the smallest member of NVIDIA's open Nemotron 3 family, which is structured around a Nano, Super, and Ultra lineup built for agentic, reasoning, and conversational workloads. NVIDIA positions the Nano tier as the most cost-efficient option for inference while still outperforming comparable models in accuracy, making it well suited to high-volume deployments where compute budgets matter. The model is distributed as open weights through a dedicated Hugging Face collection and is also accessible via NVIDIA's build.nvidia.com hosted endpoint, giving teams flexibility between self-hosting and managed serving.

The 30B-A3B designation points to NVIDIA's hybrid Mamba-Transformer mixture-of-experts architecture, which the family page highlights as a core technology enabling the efficiency claims. The backbone has proven robust enough to anchor the company's subsequent multimodal work: the Nemotron 3 Nano Omni paper describes it as the foundation for that model's expansion into native audio, image, and video understanding, paired with multimodal token-reduction techniques that lower latency and raise throughput. NVIDIA also publishes a dedicated Nano technical report alongside the broader family white paper, documenting the design choices behind this compact variant. For practitioners, the model fits naturally into agentic pipelines that need strong reasoning and tool use without paying the latency cost of a larger reasoning model.