Nemotron 3 Super - NVIDIA's open-source large model specifically designed for agent inference.

What is Nemotron 3 Super?

Nemotron 3 Super is an open-source AI model from NVIDIA with 120 billion parameters. It employs a Mamba-MoE hybrid architecture and is optimized for agent-based applications. The model supports ultra-long contexts with up to 1 million tokens, offering a 3x speedup inference and a 5x increase in throughput. It demonstrates excellent success rate on the OpenClaw task, with performance approaching that of Claude Opus 4.6. NVIDIA has also open-sourced over 10 trillion tokens of training data, a complete methodology, and 15 reinforcement learning environments, making it an ideal choice for enterprise-grade multi-agent systems.

Main features of Nemotron 3 Super

• Extremely long contextual memorySupports a context window of 1 million tokens, allowing agents to maintain a complete workflow state in complex, multi-step tasks and preventing target deviation.
• Intelligent agent task executionIt achieved a task success rate of 85.6% in intelligent agent benchmark tests such as OpenClaw, with performance approaching that of top closed-source models such as Claude Opus 4.6.
• Accelerated reasoning speedBy using multi-token prediction technology to achieve native speculative decoding, the reasoning speed is increased by 3 times, meeting the needs of real-time interaction.
• High throughput servicesThe model boasts a 5x increase in throughput compared to its predecessor, supports large-scale concurrent agent deployment, and reduces the cost of multi-agent applications.
• High-precision tool callIt enables reliable navigation within a large function library, preventing execution errors in high-risk critical environments such as cybersecurity.
• Code-based intelligent agent developmentThe model can load the entire codebase into the context at once, enabling end-to-end code generation, vulnerability patching, and automated debugging.
• Financial Analysis and ProcessingIt can load thousands of pages of reports directly into memory, saving you the trouble of repeatedly reasoning through lengthy conversations and greatly improving work efficiency.

The technical principles of Nemotron 3 Super

• Mamba-MoE hybrid architecture:The model employs an 88-layer network structure, periodically alternating between Mamba-2 layers and Transformer attention layers. The Mamba-2 layers provide linear time complexity for sequence modeling, while a small number of Transformer layers act as global anchors responsible for long-distance information routing across locations and high-precision inference, significantly improving inference throughput while maintaining powerful modeling capabilities.
• LatentMoE Implicit Hybrid Expert ArchitectureNVIDIA's pioneering new MoE design projects tokens from the hidden dimension to a smaller latent dimension before routing and expert computation. Routing and expert computation are performed in this compressed space, directly reducing parameter loading and communication by several times. The saved resources are used to increase the total number of experts and the number of experts activated, achieving the effect of "activating 4 experts at the cost of 1 expert," improving model accuracy with almost constant inference cost.
• Multi-Token Prediction AccelerationThe model predicts multiple tokens simultaneously at each position, which not only forces the model to learn multi-step causal relationships and long-term text structure to improve quality, but more importantly, it achieves native speculative decoding—the auxiliary prediction head serves as a built-in draft model to quickly generate candidate sequences, and the main model completes the verification in one forward propagation, which greatly reduces the generation latency and has minimal additional overhead.
• NVFP4 low-precision pre-trainingThe entire pre-training process was performed on the Blackwell platform using NVFP4 precision, and the 4-bit floating-point format significantly reduced the GPU memory requirements. With zero accuracy loss, the inference speed was 4 times faster than the Hopper architecture FP8, demonstrating the feasibility and efficiency of large-scale low-precision training.

Nemotron 3 Super project address

• Project official website: https://blogs.nvidia.com/blog/nemotron-3-super-agentic-ai/
• HuggingFace model libraryhttps://huggingface.co/collections/nvidia/nvidia-nemotron-v3
• Technical Papers: https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf

Application scenarios of Nemotron 3 Super

• Intelligent Agent Platform Core EngineAs the "strongest open-source model" for intelligent agent platforms such as OpenClaw, it drives multi-agent collaboration to complete complex and long-cycle tasks, solving the two major bottlenecks of context explosion and thinking tax.
• Enterprise software developmentIt empowers software development agents for companies such as CodeRabbit, Factory, and Greptile, enabling end-to-end code library generation, debugging, and vulnerability repair, achieving a SWE-Bench test score of 60.47%.
• In-depth research and analysisPowering NVIDIA AI-Q research agents, it topped the DeepResearch Bench leaderboard, supporting multi-step reasoning and information integration across massive amounts of documents.
• Network security operations and maintenanceIn high-risk environments such as autonomous and secure orchestration, high-precision tools are used to call a reliable and extensive navigation function library to prevent critical execution errors.
• Financial AnalysisIt can load thousands of pages of financial reports into memory at once and perform in-depth analysis directly without repeated reasoning, greatly improving investment research efficiency.