NVIDIA Unveils Open-Source Nemotron 3 Model Series

The global AI computing leader, NVIDIA, officially announced the open-source release of the Nemotron 3 model series on December 15, 2025. The series includes three variants: Nano, Super, and Ultra. NVIDIA officially defines this series as "the most efficient open-source model family with the best accuracy for building AI agent applications," aiming to provide developers with open-source solutions that balance performance and cost, thereby accelerating the deployment of AI agents across multiple scenarios.

According to NVIDIA's official introduction, the core competitiveness of the Nemotron 3 series stems from its innovative hybrid Mamba-Transformer Mixture-of-Experts (MoE) architecture. This architecture integrates the efficient long-sequence modeling capability of Mamba layers with the logical reasoning advantages of Transformer layers, significantly enhancing the model's adaptability and accuracy for complex tasks while reducing computational costs. This design provides the underlying technical support for scenarios involving multi-turn interactions and dynamic decision-making by AI agents.