NVIDIA has announced the launch of their latest model, the Nemotron 3 Super. This new model boasts an impressive 120 billion parameters, with 12 billion of them being active to handle complex agentic AI systems. The Nemotron 3 Super is designed to efficiently perform tasks with high accuracy for autonomous agents.
Various AI-native companies like Perplexity, CodeRabbit, Factory, Greptile, Edison Scientific, and Lila Sciences are integrating the Nemotron 3 Super model into their AI agents to enhance accuracy and reduce costs. Additionally, enterprise software platforms such as Amdocs, Palantir, Cadence, Dassault Systèmes, and Siemens are customizing the model to automate workflows in telecom, cybersecurity, semiconductor design, and manufacturing industries.
With the rise of multi-agent applications, companies face challenges like context explosion and the thinking tax. The Nemotron 3 Super addresses these issues with its 1-million-token context window, allowing agents to retain full workflow state in memory and prevent goal drift. This model has set new standards in efficiency and openness, claiming the top spot on Artificial Analysis with leading accuracy among models of similar size.
The Nemotron 3 Super utilizes a hybrid mixture-of-experts (MoE) architecture, combining Mamba layers for higher memory and compute efficiency, transformer layers for advanced reasoning, latent MoE for improved accuracy, and multi-token prediction for faster inference. Running on the NVIDIA Blackwell platform in NVFP4 precision, the model cuts memory requirements and speeds up inference without compromising accuracy.
NVIDIA is releasing the Nemotron 3 Super with open weights under a permissive license, allowing developers to deploy and customize it on workstations, data centers, or in the cloud. The model was trained on synthetic data generated using frontier reasoning models, with NVIDIA publishing the complete methodology and training datasets for researchers to fine-tune the model using the NVIDIA NeMo platform.
The Nemotron 3 Super is ideal for handling complex subtasks within multi-agent systems, enabling end-to-end code generation, financial analysis, and high-accuracy tool calling for various industries like cybersecurity. Enterprises and developers can access the model through cloud service providers, NVIDIA cloud partners, inference service providers, and data platforms and services for deployment.
To stay updated on agentic AI and NVIDIA’s latest developments, subscribe to NVIDIA AI news, join the community, and follow NVIDIA AI on social media platforms. Explore self-paced video tutorials and livestreams to learn more about the capabilities of the Nemotron 3 Super model.
For more Information, Refer to this article.
