Executive Summary
NVIDIA has launched Nemotron 3 Super, a 120-billion-parameter open model designed to power complex, multi-agent AI systems. It aims to solve the high costs and performance issues ("context explosion" and "thinking tax") inherent in agentic workflows by utilizing a hybrid mixture-of-experts (MoE) architecture and a 1-million-token context window. The model is intended for developers and enterprises building sophisticated autonomous agents, offering a more efficient and accurate solution for tasks requiring deep reasoning and long context.
Key Takeaways
* Product: NVIDIA Nemotron 3 Super, a 120-billion-parameter model with 12 billion active parameters at inference.
* Primary Function: To efficiently run complex, multi-agent AI systems that require long-form reasoning and large context, such as in software development, financial analysis, and cybersecurity.
* Key Architecture & Features:
* Hybrid MoE: Combines efficient Mamba layers with reasoning-focused Transformer layers.
* Large Context: Features a 1-million-token context window to prevent "goal drift" in long tasks.
* Performance Innovations: Uses Latent MoE and Multi-Token Prediction to boost accuracy and speed.
* Hardware Optimization: Optimized for the NVIDIA Blackwell platform, running in NVFP4 precision for up to 4x faster inference than FP8 on Hopper.
* Target Audience: AI-native companies, enterprise software platforms, developers, and researchers building agentic AI systems.
* Open Approach: Released with open weights under a permissive license, along with its full training methodology, over 10 trillion tokens of training data, and evaluation recipes.
* Availability: Available now via build.nvidia.com, Hugging Face, and a wide range of partners, including major cloud providers (Google Cloud, OCI, AWS, Azure), hardware vendors (Dell, HPE), and various inference service providers. It is packaged as an NVIDIA NIM microservice for easy deployment.
Strategic Importance
This launch positions NVIDIA to power the growing agentic AI market by providing an open, high-performance foundational model, which in turn drives demand for its Blackwell hardware. By addressing key cost and efficiency bottlenecks, NVIDIA aims to accelerate the enterprise shift from simple chatbots to more complex, autonomous AI systems.