Nemotron 3 Ultra 550B A55B is an open-weights frontier reasoning model from NVIDIA built on a hybrid Transformer-Mamba mixture-of-experts architecture, with 55B active parameters drawn from a 550B total parameter pool. This hybrid design aims to combine the long-range sequence modeling strengths of Mamba with the established representational power of Transformers, while the MoE routing allows only a fraction of the weights to activate per token, making the large total capacity more efficient at inference time. The model responds by first generating a reasoning trace and then producing a final answer, with reasoning behavior controllable through a flag in the chat template, giving developers a way to tune how much deliberation the model performs before answering.

The model is oriented toward demanding agentic and analytical workloads, including multi-step agent orchestration, coding agents, deep research, and high-accuracy reasoning over code, math, and science, where its long context window is particularly useful for sustained workflows over large documents or multi-turn task chains. Because NVIDIA publishes the weights openly on Hugging Face, teams can self-host for full control or reach the same model through managed API endpoints for quicker integration. On UnoRouter, this variant is offered as a free-tier endpoint, making it attractive for prototyping and experimentation without per-token cost, though users should be aware that free NVIDIA endpoints carry usage logging and terms specific to NVIDIA's trial service.