Nemotron 3 Ultra, Super and Nano: choosing agent models

NVIDIA's office hours session compares Nemotron 3 Ultra, Super and Nano for AI agent workflows. The speakers recommend starting with Ultra as an orchestrator and testing Super when latency matters more. They position Nano as a worker for narrowly defined tasks with a limited set of tools. These are starting points for evaluation, rather than guarantees about which model will suit a workload.

Their suggested process begins with a task-specific evaluation and a strong baseline model. Developers can then replace individual roles with smaller models, inspect failures and adjust prompts before considering fine-tuning. The discussion covers supervised fine-tuning, LoRA and reinforcement learning, including using Ultra to generate synthetic training data for Nano. The speakers describe open weights, datasets and training recipes, but give no concrete hardware requirements for a local LLM deployment.

Ultra also has a practical limitation: according to the presenters, it sometimes stops early and asks whether to continue. They suggest goal prompts, review loops or interactive pair programming to keep work moving. The session explains multi-teacher on-policy distillation as a process that combines specialized versions of Ultra into a shared checkpoint. For measuring reliability, the speakers favor complete tasks with verifiable scoring over any single reasoning or tool-use benchmark.