Start with Nano
Nano is the practical entry point for high-throughput agents and experimentation. NVIDIA lists 31.6B total parameters with roughly 3.2B active parameters (3.6B including embeddings), a hybrid Mamba-Transformer MoE design, and context support up to 1M tokens.
Evaluate Super for collaborative workloads
NVIDIA positions Super between Nano and Ultra for collaborative agents and high-volume enterprise work. Because checkpoints and details evolve, verify the current release and specifications before architecture decisions.
Use Ultra when quality is the priority
Ultra has 550B total and 55B active parameters. It is intended for deep research, complex coding, orchestration and long-running agent workflows. Most users will evaluate it through a hosted endpoint before considering direct deployment.
Decision checklist
- What are your acceptable latency and cost per request?
- Do you require self-hosting, or is a hosted endpoint acceptable?
- How long are typical prompts, tool histories and outputs?
- Does the task benefit from deeper multi-step reasoning?
- Have you reviewed the model license and provider terms?