What is OpenReasoning-Nemotron
OpenReasoning-Nemotron is a series of open-source large language models (LLMs) with powerful reasoning capabilities from NVIDIA, distilled from theDeepSeek R1 0528 model, with parameter sizes covering 1.5B, 7B, 14B, and 32B. The models focus on reasoning tasks in mathematics, science, and code, trained based on large-scale data distillation and supervised fine-tuning (SFT), setting new records on multiple benchmarks, and in mathematics, surpassingo3, demonstrating exceptional reasoning performance. The models support a "heavy" reasoning mode, further improving performance based on the GenSelect algorithm combined with the work of multiple agents.
Main Features of OpenReasoning-Nemotron
- Efficient Reasoning: Excels in mathematics, science, and code, generating high-quality reasoning solutions.
- Multiple Model Sizes: Offers models with different parameter sizes such as 1.5B, 7B, 14B, and 32B, meeting different computational resource and task requirements.
- "Heavy" Reasoning Mode: Based on the GenSelect algorithm combined with the reasoning results of multiple agents, further improving performance, with outstanding performance in mathematics and code tasks.
- Strong Baseline Models: Provides a strong starting point for future reinforcement learning (RL)-based reasoning research, helping develop more efficient reasoning technologies.
- Local Running Support: Supports 100% local running, deployed and used based on tools such as LM Studio.
Technical Principles of OpenReasoning-Nemotron
- Large-Scale Data Distillation: Uses the DeepSeek R1 0528 671B model to generate 5 million high-quality reasoning solutions, covering mathematics, science, and code. The data is used to train the OpenReasoning-Nemotron models, using data distillation to improve the models' reasoning capabilities.
- Supervised Fine-Tuning (SFT): The models are trained with supervised fine-tuning (SFT), without using reinforcement learning (RL). This demonstrates the strong potential of data distillation and provides a solid foundation for subsequent RL research.
- Multi-Agent Reasoning (GenSelect): Based on the GenSelect algorithm, the models can initiate multiple parallel reasoning processes and select the best solution from them.
- Model Architecture: Based on the Qwen 2.5 architecture, combined with data generated by the latest R1 model, ensuring the models' efficiency and accuracy on reasoning tasks.
Project Address of OpenReasoning-Nemotron
- HuggingFace Model Repository：https://huggingface.co/collections/nvidia/openreasoning-nemotron-687730dae0170059860f1f01
Application Scenarios of OpenReasoning-Nemotron
- Mathematical Problem Solving: Assists in solving complex mathematical problems in education, research, and competitions, providing detailed solution steps and reasoning support.
- Scientific Reasoning: Provides reasoning and solutions for complex problems in fields such as physics, chemistry, biology, and environmental science.
- Code Generation and Optimization: Automatically generates code snippets, optimizes code performance, and assists with code debugging, improving software development efficiency.
- Multi-Agent Collaboration: Decomposes complex tasks, selects the optimal solution through multi-agent collaboration, and optimizes system performance.
- Research and Development: Provides baseline models for reinforcement learning research, supporting the development of new technologies and exploration of reasoning algorithms.