Nvidia Nemotron 3 Super: A 120B Leap for Autonomous Agents

Nvidia has officially released Nemotron 3 Super, a 120-billion parameter reasoning model specifically architected for the next generation of autonomous AI agents. This model marks a significant technical departure from standard transformers, utilizing a hybrid Mamba state-space and Mixture-of-Experts (MoE) design.

Optimized for Agentic Reasoning

Nemotron 3 Super is designed to handle the long-context, multi-step reasoning chains required for autonomous software development and security operations. By leveraging the Mamba architecture, Nvidia claims a 5x improvement in inference speed for long sequences compared to traditional dense models, while the MoE structure ensures that the 120B model operates with the computational efficiency of a much smaller system.