Model Releases
NVIDIA Nemotron 3 Super: The Open MoE Powerhouse for Enterprise Agents
NVIDIA releases Nemotron 3 Super, a 120B MoE open-weight model designed for high-throughput agentic workloads and complex reasoning tasks.
March 11, 2026
Model ReleaseNemotron 3 Super
Introduction
On March 11, 2026, NVIDIA officially launched Nemotron 3 Super, a groundbreaking open-source AI model engineered specifically to power enterprise-grade autonomous agents. Unlike traditional monolithic architectures, this model prioritizes compute efficiency and accuracy for complex multi-agent workloads such as software development and cybersecurity triage. As the industry shifts towards agentic systems, Nemotron 3 Super represents a strategic move by NVIDIA to provide open weights that can compete with closed-source giants like OpenAI and Anthropic.
The release comes amidst a broader investment of $26 billion into open-weight AI models, signaling NVIDIA's commitment to democratizing high-performance inference. This model is not merely a chatbot but a reasoning-focused engine designed to handle autonomous tasks at scale. For developers and AI engineers, this marks a significant milestone in accessing enterprise-grade intelligence without the licensing restrictions of proprietary models.
- Released: March 11, 2026
- Provider: NVIDIA
- License: Open Weights
Key Features & Architecture
Nemotron 3 Super utilizes a Mixture of Experts (MoE) architecture, which allows the model to dynamically activate only the necessary parameters for a given task. This design choice significantly reduces computational overhead while maintaining high accuracy. The model boasts a total of 120 billion parameters, with only 12 billion active during inference, striking an optimal balance between capacity and efficiency.
Beyond parameter count, the architecture supports a massive context window essential for long-form reasoning and RAG applications. It is also specialized for agent-based AI inference, meaning it excels at breaking down complex problems into sub-tasks and executing them autonomously. The model includes advanced reasoning capabilities that efficiently complete tasks with high accuracy for autonomous systems.
- Total Parameters: 120 Billion
- Active Parameters: 12 Billion
- Architecture: Hybrid MoE
- Specialization: Agentic Systems
Performance & Benchmarks
In terms of raw capability, Nemotron 3 Super demonstrates superior performance compared to previous Nemotron iterations and key competitors. Benchmarks indicate significant improvements in MMLU (89.5%) and HumanEval (92.1%), showcasing its strength in both general knowledge and coding tasks. The model is particularly optimized for SWE-bench, where it achieves state-of-the-art results in autonomous software engineering workflows.