What is Confucius3-Math?
Confucius3-Math is a math model in the "Confucius3" series of large models open-sourced by NetEase Youdao, and is China's first open-source reasoning model focused on mathematics education. It has 14 billion parameters and is specifically optimized for K-12 mathematics education scenarios. It can run efficiently on a single consumer-grade GPU (such as RTX 4090D), with inference performance approximately 15 times that of DeepSeek R1, and a service cost as low as $0.15 per million tokens, far lower than most general-purpose large models. It performs excellently on multiple mathematical datasets, achieving a score of 98.5 in the GAOKAO-Bench (Math) evaluation framework based on college entrance examination mathematics questions, surpassing many larger general-purpose models. Through large-scale reinforcement learning and innovative algorithms, it achieves better cost-effectiveness, providing educational institutions and developers with a low-cost, high-performance AI solution.
Main Features of Confucius3-Math
-
Efficient Reasoning Capability: The model performs excellently in multiple mathematical reasoning tasks, especially achieving a score of 98.5 in the GAOKAO-Bench (Math) evaluation framework for college entrance examination mathematics questions, surpassing many larger general-purpose models.
-
Low Cost and High Performance: Through large-scale reinforcement learning and innovative algorithms, the training cost is only $26,000, and the inference performance is approximately 15 times that of DeepSeek R1. It can run efficiently on a single consumer-grade GPU (such as RTX 4090D), with a service cost as low as $0.15 per million tokens.
-
Optimization for Educational Scenarios: Specifically designed for K-12 mathematics education scenarios, effectively assisting mathematics teaching and solving problems such as unequal educational resources and insufficient personalized tutoring.
Technical Principles of Confucius3-Math
-
Large-Scale Reinforcement Learning: The model uses large-scale reinforcement learning technology, combined with the characteristics of mathematical problems, to conduct targeted training, performing excellently in mathematical reasoning tasks.
-
Innovative Algorithm Optimization: A series of innovative algorithms are adopted, significantly improving the model's inference performance and training efficiency. The inference performance is approximately 15 times that of DeepSeek R1, and the training cost is only $26,000.
-
Lightweight Design: Confucius3-Math is a lightweight model with 14 billion parameters, capable of running efficiently on a single consumer-grade GPU (such as RTX 4090D), greatly reducing hardware costs.
-
Optimization for Educational Scenarios: The model is optimized for K-12 mathematics education scenarios and can better understand and solve complex problems such as college entrance examination mathematics questions.
Project Addresses of Confucius3-Math
- Github Repository：https://github.com/netease-youdao/Confucius3-Math
- Technical Paper：https://github.com/netease-youdao/Confucius3-Math/blob/main/Confucius3-Math.pdf
- Online Demo：https://confucius.youdao.com/
Application Scenarios of Confucius3-Math
-
Personalized Mathematics Tutoring: Confucius3-Math can provide students with personalized mathematics learning plans. By analyzing students' learning progress, mastery of knowledge, and problem-solving habits, the model can generate targeted practice questions and learning suggestions, helping students better understand and master mathematical knowledge.
-
Intelligent Homework Grading and Feedback: The model can automatically grade mathematics homework and test papers, providing students with detailed solution steps and error analysis.
-
Mathematics Problem Solving and Explanation: Confucius3-Math can quickly and accurately solve various mathematical problems, providing clear solution ideas and steps.
-
Optimization and Sharing of Educational Resources: Through its open-source and low-cost characteristics, Confucius3-Math lowers the barrier for educational institutions and developers to deploy mathematics AI applications.
-
Mock Exams and Ability Assessment: The model can generate mock exam questions and assess students' abilities based on their performance.