OpenMath-Nemotron-1.5B: Advanced Mathematical Reasoning
OpenMath-Nemotron-1.5B, developed by NVIDIA, is a 1.5 billion parameter language model built upon the Qwen2.5 architecture. It has been specifically fine-tuned using the proprietary OpenMathReasoning dataset to excel in complex mathematical problem-solving.
Key Capabilities & Performance
This model demonstrates state-of-the-art performance on challenging mathematical benchmarks, including AIME24, AIME25, and HMMT-24-25. For instance, the OpenMath-Nemotron-1.5B CoT model achieves 61.6% on AIME24 and 49.5% on AIME25, with further improvements using GenSelect methods. It supports a substantial context length of up to 131,072 tokens, enabling it to process and reason through lengthy mathematical problems. The model can be utilized in three inference modes: Chain-of-Thought (CoT), Tool-Integrated Reasoning (TIR), and Generative Solution Selection (GenSelect).
Unique Aspects & Use Cases
- Specialized Mathematical Reasoning: Unlike general-purpose LLMs, this model is explicitly designed and optimized for mathematical tasks, making it highly effective for academic research and competitive math problem-solving.
- Commercial Use Ready: The model is released for commercial use under the CC-BY-4.0 license, with additional Apache License 2.0 terms.
- Reproducible Results: NVIDIA has open-sourced the entire pipeline, including code, models, and the OpenMathReasoning dataset, allowing users to fully reproduce their benchmark results and methodology.
- Optimized for NVIDIA Hardware: Designed and optimized to run efficiently on NVIDIA GPU-accelerated systems, leveraging CUDA libraries for faster inference.
Important Considerations
It is crucial to note that this model has not been instruction-tuned on general data and may not perform well outside the mathematical domain. For optimal performance in TIR or GenSelect modes, users are encouraged to utilize the reference implementation in NeMo-Skills.