Nvidia has released Nemotron 3 Super as an open-weight AI model and published 10 trillion tokens of training data to help developers build and scale autonomous AI agents more efficiently.
Nvidia has launched Nemotron 3 Super, a 120-billion-parameter open-weight AI model designed to power complex agentic AI systems. The model has 12 billion active parameters and is built to run autonomous AI agents capable of completing tasks with efficiency and high accuracy.
Released with open weights under a permissive licence, Nemotron 3 Super allows developers to deploy and customise the model on workstations, in data centres, or in the cloud. Researchers can also fine-tune the model using Nvidia’s NeMo platform, expanding its capabilities for specialised applications.
To strengthen the open ecosystem around the model, Nvidia has also published 10 trillion tokens of training data, including pre- and post-training datasets, training methodology, reinforcement learning environments, and evaluation practices. The model itself was trained entirely on synthetic data generated using frontier AI reasoning models.
Under the hood, Nemotron 3 Super uses a hybrid Mixture-of-Experts (MoE) architecture, with 120 billion parameters overall and 12 billion active during inference. A specialist activation technique allows the model to activate four expert specialists for the cost of one, while Multi-Token Prediction enables the model to predict multiple future words simultaneously, delivering three times faster inference.
Nvidia claims the model can deliver up to five times higher throughput and up to two times higher accuracy, along with up to four times faster inference compared with FP8 when running on Blackwell GPUs using NVFP4 precision. The model also features a 1-million-token context window, which, according to Nvidia, means “Nemotron 3 Super has a 1-million-token context window, allowing agents to retain full workflow state in memory and preventing goal drift.”
Designed for multi-agent systems, the model can load entire codebases into context, process thousands of pages of reports, and execute complex reasoning tasks, while enabling high-accuracy tool calling for applications such as cybersecurity.
Nemotron 3 Super also topped the Artificial Analysis evaluation for model efficiency and openness and powered an AI research agent that secured the No.1 ranking on DeepResearch Bench and DeepResearch Bench II leaderboards.
The model is available through Perplexity, OpenRouter, Hugging Face, and build.nvidia.com, and is already being integrated into AI coding agents such as CodeRabbit, Factory, and Greptile.
