New open-source math model Light-R1-32B surpasses equivalent DeepSeek performance with only $1000 in training costs

Researchers have launched Light-R1-32B, a new open source AI model to solve advanced math problems. Now available Hug the face Under the permitted Apache 2.0 license, businesses and researchers can make, deploy, adjust or modify, or even for commercial purposes as they wish.

32 billion parameters (model settings) models exceed the performance of similar size (or even larger) open source models, such as DeepSeek-r1-Distill-dillama-70B and DeepSeek-r1-distill-distill-qwen-32b on third-party third-party American Invitational Mathematics Exam (AIME) The benchmark has 15 math problems designed for extreme senior students, with the allocated time limit of 3 hours.

Developed by Liang Wen, Fenrui Xiao, Xin He, Yunke Cai, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia and Xiangzheng Zhang, the model surpasses previous open-source alternatives on competitive math benchmarks.

Incredibly, the researchers completed training on the model with 12 NVIDIA H800 GPUs, with an estimated total cost of $1,000. This makes Light-R1-32B one of the easiest and practical ways to develop high-performance mathematical-specific AI models. However, it is important to remember that the model is trained in a variant of Alibaba’s open source QWEN 2.5-32B – Classroomthink that this itself has higher preliminary training costs.

In addition to the models, the team has released its training datasets, scripting and evaluation tools, providing a transparent and accessible framework for building math-centric AI models.

The arrival of Light-R1-32b follows similar efforts from competitors, e.g. Microsoft Orca-Math.

A new math king appears

To help Light-R1-32B solve complex mathematical reasoning, researchers were trained on models that were not equipped with long chain (COT) reasoning. They use course-based supervised fine-tuning (SFT) and direct preference rate limiting (DPO) to refine problem-solving capabilities.

When evaluated, Light-R1-32b reached 76.6 on 64.6 on AIME24 and AIME25, surpassing DeepSeek-R1-Distill-Qwen-32b with scores of 72.6 and 54.9, respectively.

This improvement shows that curriculum-based training methods can effectively enhance mathematical reasoning even in the initial lack of training for long-bed models.

Fair benchmarking

To ensure fair benchmarking, researchers purified the training data on common inference benchmarks including AIME24/25, MATH-500 and GPQA DIAMOND, thus preventing data leakage.

They also implemented difficulty-based response filtering using DeepScaler-1.5b-preiview, eventually forming 76,000 sample datasets in the first phase of supervised fine-tuning. The second more challenging dataset of 3,000 examples further improves performance.

After training, the team merged multiple trained Light-R1-32Bs, resulting in additional benefits. It is worth noting that although the model is specialized in mathematics, it maintains strong generalization skills for scientific reasoning tasks (GPQA).

How businesses benefit

Light-R1-32B is released under Apache License 2.0 (allowed open source license), allowing free use, modification and commercial deployment without the need for open source derivative works. This is an attractive option for businesses, AI developers and software engineers who want to integrate or customize proprietary application models.

The license also includes a royalty-free, global patent grant that reduces legal risks to businesses while discouraging patent disputes. Companies can freely deploy Light-R1-32B in commercial products, thus fully controlling their innovation while benefiting from open and transparent Artificial Intelligence Ecosystem.

For CEOs, CTOs and their leaders, Apache 2.0 ensures cost efficiency and supplier independence, eliminates licensing fees and restrictive dependencies on proprietary AI Solutions. AI developers and engineers have the flexibility to fine-tune, integrate and scale models without limitations, which is ideal for dedicated mathematical reasoning, research and enterprise AI applications.

However, because the license does not provide any warranty or liability coverage, organizations should conduct their own security, compliance and performance evaluations in critical environments when deploying Light-R1-32B.

Transparency of low-cost training and optimization for mathematical problem solving

The researchers stress that Light-R1-32B provides a proven, cost-effective way to train powerful long COT models in a dedicated domain.

By sharing their approaches, training data and code, they aim to reduce cost barriers to enable high-performance AI development. Going forward, they plan to explore reinforcement learning (RL) to further enhance the model’s inference ability.