NVIDIA Launches Nemotron 3 Super Model for Agentic AI Systems

NVIDIA has introduced Nemotron 3 Super, a 120-billion-parameter open model built to enhance large-scale agentic AI systems. The hybrid mixture-of-experts model is optimized for the NVIDIA Blackwell platform and delivers up to five times higher throughput and double the accuracy compared to the previous Nemotron Super version.

Nemotron 3 Super features a 1-million-token context window, enabling autonomous agents to maintain full workflow memory and avoid goal drift during complex tasks. The model activates 12 billion parameters at inference and uses a combination of Mamba and transformer layers for improved efficiency and reasoning. It also incorporates multi-token prediction to accelerate inference by up to three times.

The model’s open weights and training data are released under a permissive license, allowing developers to deploy and customize it across workstations, data centers, and cloud environments. It was trained on more than 10 trillion tokens and includes methodologies for reinforcement learning and evaluation.

Nemotron 3 Super is being integrated by companies such as Perplexity, CodeRabbit, and Factory for AI search and software development agents. Enterprises including Amdocs, Palantir, and Siemens are adopting it to automate workflows in telecommunications, cybersecurity, and manufacturing. The model is available through build.nvidia.com, Perplexity, OpenRouter, and Hugging Face, with deployment options across major cloud providers and NVIDIA partners.