What Is Nemotron 3 Super
Nemotron 3 Super is a 120-billion-parameter open-source AI model released by Nvidia. It uses a Mamba-MoE hybrid architecture and is optimized for agentic applications. The model supports an ultra-long context of 1 million tokens, delivers 3x faster inference, and provides 5x higher throughput. InOpenClawtask success rate performance is excellent, with performance approachingClaude Opus 4.6. Nvidia has also open-sourced over 10 trillion tokens of training data, the complete methodology, and 15 reinforcement learning environments, making it an ideal choice for enterprise-grade multi-agent systems.
Main Features of Nemotron 3 Super
- Ultra-Long Context Memory: Supports a 1-million-token context window, enabling agents to maintain complete workflow state in complex multi-step tasks and preventing goal drift.
- Agentic Task Execution: Achieves an 85.6% task success rate on agentic benchmarks such as OpenClaw, with performance approaching top closed-source models such as Claude Opus 4.6.
- Accelerated Inference Speed: Uses multi-token prediction technology to achieve native speculative decoding, increasing inference speed by 3x and meeting real-time interaction requirements.
- High-Throughput Serving: The model delivers 5x higher throughput than the previous generation, supports large-scale concurrent agent deployment, and reduces the cost of multi-agent applications.
- High-Precision Tool Calling: Can reliably navigate and operate within large function libraries, preventing execution errors in high-risk critical environments such as cybersecurity.
- Code Agent Development: The model can load an entire codebase into context at once, enabling end-to-end code generation, vulnerability fixing, and automated debugging.
- Financial Analysis Processing: Can load thousands of pages of reports directly into memory, eliminating the trouble of repeatedly re-reasoning in lengthy conversations and greatly improving work efficiency.
Technical Principles of Nemotron 3 Super
- Mamba-MoE Hybrid Architecture:The model uses an 88-layer network structure, periodically alternating Mamba-2 layers and Transformer attention layers. The Mamba-2 layers provide sequence modeling efficiency with linear time complexity, while a small number of Transformer layers serve as global anchors responsible for long-distance information routing across positions and high-precision reasoning, significantly improving inference throughput while maintaining strong modeling capability.
- LatentMoE Latent Mixture-of-Experts Architecture: A new MoE design pioneered by Nvidia that first 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 volume by several times. The saved resources are used to increase the total number of experts and the number of activated experts, achieving the effect of 'activating 4 experts at the cost of 1 expert,' improving model accuracy with almost unchanged inference cost.
- Multi-Token Prediction Acceleration: The model predicts multiple future tokens simultaneously at each position. This not only forces the model to learn multi-step causal relationships and long-range text structure to improve quality, but more importantly enables native speculative decoding - the auxiliary prediction head acts as a built-in draft model to quickly generate candidate sequences, and the main model completes verification in a single forward pass, greatly reducing generation latency with minimal additional overhead.
- NVFP4 Low-Precision Pretraining: Pretraining is performed entirely on the Blackwell platform using NVFP4 precision. The 4-bit floating-point format greatly reduces memory requirements. With zero accuracy loss, inference speed is 4x faster than FP8 on the Hopper architecture, demonstrating the feasibility and efficiency of large-scale low-precision training.
Nemotron 3 Super Project Address
- Project Website：https://blogs.nvidia.com/blog/nemotron-3-super-agentic-ai/
- HuggingFace Model Repository：https://huggingface.co/collections/nvidia/nvidia-nemotron-v3
- Technical Paper：https://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Super-Technical-Report.pdf
Application Scenarios of Nemotron 3 Super
- Core Engine of Agent Platforms: As the "strongest open-source model" for agent platforms such as OpenClaw, it drives multi-agent collaboration to complete complex long-horizon tasks, solving the two major bottlenecks of context explosion and thinking tax.
- Enterprise-grade Software Development: Empowers software development agents at companies such as CodeRabbit, Factory, and Greptile to achieve end-to-end generation, debugging, and vulnerability fixing at the codebase level, reaching 60.47% on the SWE-Bench test.
- Deep Research and Analysis: Drives the NVIDIA AI-Q research agent, which won first place on the DeepResearch Bench leaderboard, supporting multi-step reasoning and information integration across massive document collections.
- Cybersecurity Operations: In high-risk environments such as autonomous security orchestration, it reliably navigates vast function libraries through high-precision tool calling, preventing critical execution errors.
- Financial Analysis: Loads thousands of pages of financial reports into memory at once for direct in-depth analysis, without the need for repeated re-reasoning, greatly improving investment research efficiency.