Model family
Use this comparison as a starting point, then verify deployment requirements, licenses and current versions on the linked official pages.
High-throughput agents
The smallest member of the family, designed to pair strong reasoning with inference efficiency. NVIDIA describes a hybrid Mamba-Transformer mixture-of-experts architecture.
Collaborative agents
Positioned by NVIDIA for collaborative agents and high-volume workloads. Confirm current checkpoints and specifications on the official family page.
Frontier reasoning
The largest Nemotron 3 model, built for orchestration, coding agents, deep research and complex multi-step workflows.
Selection guide
Choose Nano when throughput, cost control, and high call volume matter. It is the clearest starting point for local or self-hosted experimentation where supported hardware is available.
Evaluate Super for collaborative-agent and enterprise-volume scenarios, while checking the official page for the latest release status and exact checkpoint details.
Choose Ultra for demanding reasoning and orchestration where model quality is the priority and you can use a hosted endpoint or appropriate infrastructure.
See official access options