NVIDIA-Nemotron-Nano-9B-v2

Architecture Overview:
• Hybrid Architecture: Primarily Mamba-2 and MLP layers combined with just four Attention layers (Nemotron-H design)
• Context Window: Supports up to 128K tokens for extended document processing
• Model Size: 9 billion parameters trained from scratch by NVIDIA
• Training Period: June 2025 - August 2025 with data cutoff of September 2024
• Training Infrastructure: Built using Megatron-LM and NeMo-RL frameworks

Training Methodology:
• Pretraining Corpus: Over 20 trillion tokens of high-quality curated and synthetically-generated data
• Data Sources: English Common Crawl (3.36T tokens), Multilingual Common Crawl (812.7B tokens), GitHub Crawl (747.4B tokens)
• Synthetic Data: Leverages reasoning traces from DeepSeek R1, Qwen3-235B-A22B, Nemotron 4 340B, and other state-of-the-art models
• Domain Coverage: Code (43 programming languages), legal, math, science, finance, multilingual text (15 languages)
• Post-Training: Specialized reasoning-focused instruction tuning with synthetic reasoning traces

Performance Characteristics:
• Reasoning Benchmarks: 72.1% AIME25, 97.8% MATH500, 64.0% GPQA, 71.1% LiveCodeBench
• Instruction Following: 90.3% IFEval (Instruction Strict), 66.9% BFCL v3 for function calling
• Long Context: 78.9% RULER at 128K context length
• Reasoning Modes: Supports both "reasoning-on" (with  tags) and "reasoning-off" modes via system prompts
• Runtime Control: Unique thinking budget control allows specification of maximum reasoning tokens
• Multilingual: Supports English, German, Spanish, French, Italian, and Japanese with quality improvements from Qwen integration

‍

Primary Use Cases:
• Mathematical Reasoning: Exceptional performance on AIME, MATH500, and competition-level problems
• Scientific Analysis: Strong GPQA scores for graduate-level science questions
• Code Generation: 71.1% LiveCodeBench with support for 43 programming languages
• AI Agent Systems: Controllable reasoning makes it ideal for multi-step agent workflows
• Customer Support: Reasoning budget control enables balance between accuracy and response time
• Function Calling: Native tool-calling support via BFCL v3 benchmark validation

Enterprise Applications:
• RAG Systems: 128K context window supports extensive document retrieval and analysis
• Chatbots: Multilingual support (6 languages) for global customer engagement
• Content Moderation: Trained with Nemotron Content Safety Dataset V2 for safe outputs
• Educational Tools: Mathematical and scientific reasoning capabilities for tutoring applications
• Research Assistance: Long-context support for analyzing papers, reports, and technical documents

Edge & Latency-Sensitive Deployments:
• Efficient Architecture: Mamba2-Transformer hybrid enables faster inference than pure attention models
• Variable Compute: Thinking budget control optimizes for time-critical applications
• Streaming Support: Compatible with vLLM streaming for real-time response generation
• Hardware Optimization: Optimized for NVIDIA GPUs (A10G, A100, H100) with TensorRT-LLM support

‍