Nvidia Unveils Nemotron 3 Super Open-Source AI Model for Agentic AI Systems

Nvidia also released a new open-source artificial intelligence (AI) designed for the implementation of complex agentic workflows, which was developed by NVidia. The Nemotron 3 Super is a hybrid mixture-of-experts (MoE) model, which has been described as complete tasks with high accuracy for autonomous agents and purifies the work of advanced reasoning capabilities. Several AI companies, including Perplexity, already have deployed the new model for its new agentic Computer platform. Additionally, it is also being hosted on public repositories to allow interested people to download and run the model locally.

Nvidia’s Nemotron 3 Super AI Model Released

tech giant, spokesman and detailed the new open-source AI model in a blog post. Nvidia’s website, Hugging Face platform, Perplexity and OpenRouter is currently hosting the Nemotron 3 Super as part of the family’ nemoticon 3 (Nirotor 3) super. Furthermore, it is also on sale at the Dell Enterprise Hub and is adapted to be used for on-premise deployment in its Dell AI Factory.

In the new model, it solves the problem of context and the cost of reasoning is increased. For example, AI models designed for agentic workflows tend to generate more tokens (additional interaction of each agent or sub-agent requires sending the full context) and thus generating an increase in token numbers. Similarly, multi-level thinking is also important for performing complex tasks and can significantly increase the cost of running the model.

The Nemotron 3 Super is a hybrid architecture with 120 billion parameters and 12 billion active parameters. It also receives a context window of one million tokens, which allows agents to retain full workflow memory. Additionally, its development also used a technique called Latent MoE that increases accuracy by activating four experts for the cost of one to generate the next token at inference.

The tech giant said it is’releasing the open-source model with open weights under a permissive licence’. Using frontier reasoning models, the company says that Nemotron 3 Super was trained on synthetic data generated on the dataset and training. Nvidia said the full methodology is being published by it, including over 10 trillion tokens or pre and post-training datasets, 15 training environments for reinforcement learning and evaluation recipes.