Nemotron 3 Ultra 550B A55B (Reasoning)

NVIDIA Nemotron 3 Ultra 550B A55B (Reasoning) is a frontier-scale large language model designed specifically for complex agentic workflows, long-context analysis, and advanced multi-step reasoning. Featuring 550 billion total parameters with 55 billion active parameters per token, it is the largest and most capable model in the Nemotron 3 family. The model is optimized for high-stakes analytical workloads across diverse domains including coding, mathematics, and scientific research.

The model utilizes a sophisticated Latent Mixture-of-Experts (LatentMoE) architecture that interleaves Mamba-2 layers with transformer-based Attention layers. By projecting tokens into a lower-rank latent space before routing them to experts, the architecture enables the use of significantly more expert specialists while maintaining the inference cost of a smaller model. Additionally, the inclusion of Multi-Token Prediction (MTP) layers facilitates faster text generation and improved coherence by predicting multiple future tokens in a single forward pass.

A key feature of Nemotron 3 Ultra is its native Reasoning Mode, which allows the model to generate an internal thinking trace before providing a final response. This trace, often enclosed within <think> tags, helps the model navigate complex logic and planning tasks. The model supports a massive context window of up to 1 million tokens, leveraging the linear-time complexity of Mamba-2 to maintain performance over extremely long sequences, making it suitable for analyzing entire codebases or dense technical documents.

Post-trained using an advanced pipeline that includes Supervised Fine-Tuning (SFT), Reinforcement Learning from Human Feedback (RLHF), and Multi-teacher On-Policy Distillation (MOPD), Nemotron 3 Ultra is localized for high-accuracy performance in multiple languages. It supports English, French, Spanish, Italian, German, Japanese, Korean, Hindi, Brazilian Portuguese, and Chinese.