NVIDIA Nemotron 3 Ultra is an open large language model with 550 billion total parameters and 55 billion active parameters. It is built on a hybrid Transformer-Mamba Mixture-of-Experts (MoE) architecture, designed to deliver frontier intelligence at a fraction of the compute cost of dense models of equivalent quality.

Agents don’t just answer once. They plan, call tools, delegate work to sub-agents, check results, and keep going across hundreds of turns. Every step adds tokens and compute, so the metrics that matter are task completion at useful accuracy, time-to-finish, and cost-per-task.

Nemotron 3 Ultra addresses this directly. Its MoE architecture activates only 55B of its 550B parameters per forward pass, keeping throughput high even at million-token context lengths. This means agents can sustain planning, tool calling, and self-correction loops that span hundreds of turns while helping maintain coherence and manage cost.

Nemotron 3 Ultra excels in workloads that require sustained multi-step reasoning: