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

Researchers applied a light-R1-32B with a new open source AI model optimized to solve advanced math problems. Already available Hug face Under the permissible Apache 2.0 license – free, placement for enterprises and researchers, is free for commercial purposes, as they want to be subtle adjustment or desired.

32 billion parameter (number of model parameters) Model, DEEPSEK-R1 Distill-LLA-70B-70B and DEEPSEK-R1-DISTREL-70B-32B, exceeds the performance of the same size (and larger) open source models American invitation math exam (aime) Benchmark with 15 math problems designed for extremely advanced students and the time limit of time allocated within 3 hours.

Liang Wen, Fenrui Xiao, Xin He, Yeke Cai, Yunke Cai, Qi, Yongcheng Zou, Yonghag Jiagzheng Zhang, previous open source alternatives about competitive math prices

Incredible, researchers, 12 hours have completed a $ 7,000 model preparation in 12 NVIDIA H800 GPU. This is one of the most favorable and practical approaches to develop a high-performance of high-performance mathematics, lighth-R1-32B. However, it is important to remember that the model is taught to the option of the Alibaba’s open source Qwen 2.5-32b instructionsIt is supposed to have higher level training costs.

Along with the model, the team has issued training information and scripts and evaluation tools and presented a transparent and accessible framework to establish math focused AI models.

The coming of the light-r1-32b, for example, follows similar efforts to rivals Microsoft Orca-Mathematics.

The king of a new mathematics emerges

To help light-R1-32B to help complex mathematics, researchers have trained on a model that is not an unpredictable model of the long chain. To specify the problem of solving the problem, they applied a delicate arrangement (SFT) and directly preferred Otomation (DPO) based on the curriculum.

When evaluated, AIME24 and 64.6 in AIME24 and 64.6 in 76.6 in 76.6 in 76.6 and reached 76.6 in AIME24 and 64.6, respectively 72.6 and 54.9.

This improvement increases the mathematical justification when training the curriculum-based training approach is originally training from models that do not have a long time.

Fair price

To ensure fair benchking, researchers prevent information from leaking in general thinking-500 and GPGA diamonds, including aime24 / 25, math-500 and GPGA diamonds.

Deepscscaler-1.5B-Preview also applied to the reply-based response filter, resulting in 76,000 instances for the first phase of controlled subtle adjustment. The second, more challenging performance is even more difficult in a more challenging version of 3000 samples.

After the training, the team combined numerous trained versions of a light-R1-32B that caused additional earnings. It should be noted that despite the model of mathematics, the model protects strong generalization skills related to scientific justification tasks (GPGA).

How can enterprises benefit

Light-R1-32B, without requiring an open source function that allows you to operate a Perissive License 2.0, Perisive, Modification and Commercial Placement, Modification and Commercial Placement. It makes businesses that want to model or customize for property applications, make an attractive choice for AI developers and software.

The license also includes reducing the legal risks of the enterprises, which is discouraged in the world-class patent grant, a world-class patent grant, patent disputes. While companies benefit from open and transparency, protecting full control over updates, they can freely place a light-R1-32B in commercial products EU ecosystem.

For CEOS, CTOs and IT leaders, Apache 2.0 provides expense efficiency and seller’s independence of licensing fees and restrictive dependence on property AI solutions. AI developers and engineers are preparing for subtle, integration and expansion of the model, making subtle regulation, specialized mathematics, research and enterprise AI applications.

However, as licensed guarantees or liabilities are not ensured, organizations must conduct their safety and performance assessments before lighting their own safety, compliance and performance assessments in critical environments.

Transparency in the field of low-precious training and optimization for solving the problem of mathematics

Researchers emphasize that the light-R1-32B is approved to bring up strong long cot models in specialized domains.

By sharing the methodology, training information and code, they aim to reduce cost barriers for high-performance AI development. Looking forward, plans to explore the reinforcement learning (RL) to further enhance the fact that the model is further increasing the basis.