Confucius3-Math - NetEaseYouDao Launches Open Source Reasoning Model Focused on Mathematics Education

What is Confucius 3-Math?

Confucius3-Math is the first domestic open source reasoning model focused on mathematics education open-sourced by NetEaseYouDao. With 14 billion parameters, optimized for K-12 math education scenarios, it can run efficiently on a single consumer GPU (such as RTX 4090D), with an inference performance of about 15 times that of DeepSeek R1, and a cost of service as low as $0.15 per million tokens, which is much lower than that of most general-purpose large models. It performs well on multiple math datasets, scoring up to 98.5 points in the GAOKAO-Bench (Math) evaluation framework based on college entrance exam math questions, outperforming many general-purpose models with much larger scales.

Key Features of Confucius3-Math

• Highly effective reasoning skills: The model performs well on several mathematical reasoning tasks, especially scoring 98.5 on the GAOKAO-Bench (Math), a high-stakes mathematics assessment framework, outperforming many larger general-purpose models.
• Low cost and high performance: With large-scale augmented learning and innovative algorithms, training costs only $26,000 and inference performance of about DeepSeek R1 15x. Runs efficiently on a single consumer GPU, such as the RTX 4090D, at a cost per million of service. token As low as $0.15.
• Education Scene OptimizationIt is designed for K-12 math education scenarios, effectively assisting math teaching and solving the problems of unequal educational resources and insufficient personalized tutoring.

Confucius3-Math's project address

• Github repository:: https://github.com/netease-youdao/Confucius3-Math
• Technical Papers:: https://github.com/netease-youdao/Confucius3-Math/blob/main/Confucius3-Math.pdf
• Online Demo:: https://confucius.youdao.com/

How to use Confucius3-Math

• Online Experience: Visit the online Demo address, where users can quickly experience the model's functionality by entering math problems directly into the web page and viewing the model-generated answers and solution steps.
• Download and use: Visit Confucius3-Math's GitHub repository to download the model files and related code. According to the documentation in the repository, users can install the necessary dependencies and run the model in their local environment for more in-depth testing and research.
• developer integration: Developers can integrate Confucius3-Math into their own educational applications or platforms. For example, develop a math learning app that uses the model to provide students with personalized learning plans; or add intelligent homework correction functions to an online education platform.

Core Benefits of Confucius3-Math

• Excellent inference performance: The reasoning performance of Confucius3-Math is about DeepSeek R1 15 times. Excellent performance on several math datasets (e.g., GAOKAO-Bench (Math), MathBench (K12), etc.), especially on the High School Math Question Assessment Framework, with a score of 98.5.
• Significant cost-effectivenessThe cost of training is only $26,000, and the service cost is as low as $0.15 per million tokens, much lower than most general-purpose large models, which greatly reduces the threshold of AI application in education scenarios.
• Lightweight design: With 14 billion parameters, the model is a lightweight model that runs efficiently on a single consumer-grade GPU, such as the RTX 4090D, without the need for costly hardware investments.
• Targeted optimization: Confucius3-Math is designed for K-12 math education scenarios to better understand and solve complex problems such as high-stakes math problems, and to effectively assist in math instruction.
• Large Scale Augmented Learning: Targeted training of the model through large-scale augmented learning techniques that incorporate the characteristics of the mathematical problem significantly improves the model's inference performance.

People for whom Confucius3-Math is suitable

• K-12 students: is the primary target user group for Confucius3-Math. The model focuses on mathematics education and can provide personalized learning support for students from elementary to high school.
• Primary and secondary math teachers: Confucius3-Math can be used to improve the efficiency and quality of teaching, for example:Automatically correct students' math homework to reduce teachers' workload.
• Offline Training InstitutionsConfucius3-Math can be integrated into the teaching system to provide students with richer learning resources and personalized tutoring services.
• Educational researchers: Models can be used to conduct educational research, analyze students' learning behaviors and learning effects, and provide data support for educational reform and teaching method innovation.
• head of a household: Parents can use the features of Confucius3-Math to help their children solve problems in math learning, provide learning suggestions, and better tutor their children in math learning. The model-generated learning resources and practice problems provide more learning materials to support children's independent learning.