NVIDIA Launches Nemotron 3 Super with 5x Higher Throughput

Today, NVIDIA introduced the Nemotron 3 Super, a model featuring 120 billion parameters, with 12 billion active parameters designed for complex agentic AI systems. This model allows for high-accuracy task completion and is already being integrated by companies like Perplexity for search and software development, significantly improving precision while reducing costs.

Industry leaders such as Amdocs and Palantir are utilizing the model to automate workflows in sectors like telecommunications and cybersecurity. However, companies face two main challenges: context explosion and thinking tax. Nemotron 3 Super addresses these issues with a 1-million-token context window, enabling agents to retain full workflow state in memory.

The model also showcases impressive results in efficiency and openness benchmarks, claiming top positions in tests measuring AI's ability to conduct multistep research. Its architecture combines several innovations, delivering up to 5x higher throughput and up to 2x higher accuracy than its predecessor.

NVIDIA is releasing Nemotron 3 Super with open weights, allowing developers to customize the model for deployment in the cloud or on-premises. The model was trained on synthetic data, and NVIDIA is publishing the complete training methodology, including over 10 trillion tokens of data.

Nemotron 3 Super is designed to handle complex subtasks within multi-agent systems, significantly enhancing efficiency in areas like financial analysis and cybersecurity automation. The model is accessible through various cloud services and partners, such as Google Cloud and Amazon Web Services, simplifying its integration into business processes.