Nemotron 3 Super Released
NVIDIA's new open model uses a hybrid Mamba-Transformer architecture to power AI agents.
NVIDIA has launched Nemotron 3 Super, a groundbreaking open-source model designed specifically for powering the next generation of AI agents. The model features a massive 120 billion parameter Mixture of Experts (MoE) architecture, but only activates 12 billion parameters during inference, making it surprisingly efficient for its size. Its key innovation is a hybrid design that combines Mamba (a state-space model) with a Transformer, aiming to capture both long-range dependencies and efficient sequence processing crucial for complex reasoning.
This architecture is purpose-built for 'agentic reasoning'—the ability for an AI to break down complex problems, plan multi-step actions, and execute workflows. NVIDIA positions Nemotron 3 Super as a foundational model for developers building sophisticated AI assistants, coding agents, or research tools that require advanced logic and planning. By releasing it as an open model, NVIDIA is providing a powerful, customizable alternative to closed models for the developer community working on autonomous AI systems.
- Uses a hybrid Mamba-Transformer MoE architecture with 120B total and 12B active parameters.
- Specifically engineered for 'agentic reasoning' to plan and execute multi-step tasks.
- Released as an open-source model, providing a powerful alternative for building advanced AI agents.
Why It Matters
Provides an open, high-performance foundation for developers building complex, reasoning-based AI agents and autonomous systems.