OpenReasoning-Nemotron - A series of open-source inference models from NVIDIA
OpenReasoning-Nemotron is a series of powerful large language models (LLMs) open sourced by NVIDIA. It is distilled from the DeepSeek R1 0528 model and has parameter sizes of 1.5B, 7B, 14B and 32B.
What is OpenReasoning-Nemotron?
OpenReasoning-Nemotron is a series of powerful large-scale language models (LLMs) from NVIDIA, distilled from the DeepSeek R1 0528 model, with parameter sizes ranging from 1.5B, 7B, 14B, and 32B. The model focuses on reasoning tasks in mathematics, science, and coding domains. Based on large-scale data distillation and supervised fine-tuning (SFT) training, it has set new records in multiple benchmarks, surpassing o3 mathematically and demonstrating outstanding reasoning performance. The model supports "heavy" reasoning modes, further enhancing performance by combining the work of multiple agents using the GenSelect algorithm.
Main functions of OpenReasoning-Nemotron
- Efficient ReasoningIt excels in fields such as mathematics, science, and coding, and can generate high-quality reasoning solutions.
- Multi-model scaleIt provides models with different parameter scales, such as 1.5B, 7B, 14B and 32B, to meet different computing resource and task requirements.
- "Heavy" reasoning modeBased on the GenSelect algorithm and combining the inference results of multiple agents, performance is further improved, and it performs outstandingly in mathematical and coding tasks.
- Strong baseline modelThis provides a strong starting point for future research on reinforcement learning (RL)-based reasoning, and helps to develop more efficient reasoning techniques.
- Local operation supportIt supports 100% local operation and deployment and use based on tools such as LM Studio.
The technical principles of OpenReasoning-Nemotron
- Large-scale data distillationGenerates 5 million high-quality inference solutions using the DeepSeek R1 0528 671B model, covering mathematical, scientific, and coding domains. The data is used to train the OpenReasoning-Nemotron model, and data distillation is used to improve the model's inference capabilities.
- Supervised fine-tuning (SFT)The model was trained using supervised fine-tuning (SFT) without reinforcement learning (RL). This demonstrates the powerful potential of data distillation and provides a solid foundation for future RL research.
- Multi-agent reasoning (GenSelect)Based on the GenSelect algorithm, the model can initiate multiple parallel inference processes and select the best solution from them.
- Model ArchitectureBased on the Qwen 2.5 architecture and combined with data generated from the latest R1 model, the efficiency and accuracy of the model in inference tasks are ensured.
OpenReasoning-Nemotron project address
- HuggingFace model library: https://huggingface.co/collections/nvidia/openreasoning-nemotron-687730dae0170059860f1f01
Application scenarios of OpenReasoning-Nemotron
- Mathematical Problem SolvingIt assists in solving complex mathematical problems in education, scientific research, and competitions by providing detailed solution steps and reasoning support.
- Scientific reasoningIt provides reasoning and solutions for complex problems in fields such as physics, chemistry, biology, and environmental science.
- Code generation and optimizationIt automatically generates code snippets, optimizes code performance, and assists in code debugging, thereby improving software development efficiency.
- Multi-agent collaborationDecompose complex tasks, select the optimal solution through multi-agent collaboration, and optimize system performance.
- Research and DevelopmentIt provides a baseline model for reinforcement learning research, supporting the development of new technologies and the exploration of inference algorithms.