NVIDIA Nemotron 3 Super

Architecture Overview:
• Hybrid Mixture of Experts (MoE) with Mamba-transformer architecture
• 120B total parameters with 12B activated per forward pass via sparse MoE routing
• Optimized for running many collaborating agents per application on single GPU
• Hybrid Mamba-transformer delivers significantly higher token generation throughput vs pure transformer
• Thinking budget optimization avoids overthinking and ensures predictable inference costs
• 1M token context length for processing extensive codebases and long-horizon workflows
• Single GPU deployment: 1×B200, 1×GB200, 2×H100, 1×H200, 4×A100, 4×L40S, 1×DGX Spark, 1×RTX 6000

Training Methodology:
• Trained with NVIDIA-curated high-quality synthetic data from expert reasoning models
• Reinforcement learning alignment to reason like humans across various categories
• Open data: fully transparent synthetic dataset generated using frontier open reasoning models
• Open recipes: NVIDIA development techniques and tools for customization and optimization
• Post-training optimizations for powerful, transparent, and adaptable deployment

Performance Characteristics:
• Leading accuracy across GPQA Diamond, AIME 2025, LiveCodeBench, IFBench, BFCL benchmarks
• Highest compute efficiency via hybrid Mamba-transformer architecture
• MoE architecture reduces compute and meets stringent latency requirements
• Thinking budget optimizes for lower, predictable inference cost
• Multi-agent optimization: high accuracy for reasoning, tool calling, instruction following
• Significantly higher token generation throughput enabling faster thinking

‍

Multi-Agent Software Development:
• Optimized for running many collaborating agents per application on single GPU
• Code summarization, generation, refactoring across multiple agent workflows
• Leading accuracy on LiveCodeBench for competitive programming tasks
• High accuracy for tool calling and instruction following in complex coding workflows
• 1M context supporting entire codebases and long-horizon development tasks

Financial Services Automation:
• Accelerate loan processing by extracting data and analyzing income patterns
• Detect fraudulent operations reducing cycle times and risk
• Multi-agent workflows for comprehensive financial analysis
• Thinking budget ensures predictable costs for high-volume operations

Cybersecurity Operations:
• Automatically triage vulnerabilities with multi-agent coordination
• Perform in-depth malware analysis across security tools
• Proactively hunt for security threats with agentic workflows
• High accuracy for instruction following in security-critical operations

Search & Productivity Agents:
• Leading accuracy on IFBench and BFCL for instruction following and function calling
• Multi-agent search workflows to increase productivity
• Thinking budget optimization for cost-effective at-scale deployment
• 1M context for processing extensive research materials

Retail Optimization:
• Optimize inventory management with multi-agent coordination
• Real-time personalized product recommendations and support
• Enhance in-store service with collaborative agent systems
• Predictable inference costs via thinking budget optimization

Open-Source Customization:
• Open weights: NVIDIA license for enterprise flexibility and data control
• Open data: fully transparent NVIDIA-generated synthetic training data
• Open recipes: development techniques for building custom reasoning models
• Deploy anywhere: laptop to cloud via NVIDIA NIM
• Full transparency and adaptability for researchers and enterprises

‍