Youdao 14B Low-Cost Lightweight Model 'Zi Yue 3' Open Source, Outperforming Larger Models in Mathematical Reasoning}
On June 23, NetEase Youdao announced the open source of 'Zi Yue 3' series large models, focusing on math education. It offers high performance at low cost, surpassing larger models in mathematical reasoning tasks.
On June 23, NetEase Youdao officially announced the open source of the mathematical model series 'Zi Yue 3' (English name: Confucius3-Math). This is the first open source inference model in China dedicated to math education that can efficiently run on a single consumer-grade GPU. It has achieved state-of-the-art performance in a series of mathematical reasoning tasks, surpassing many larger general models. The open sourcing of this model provides a low-cost, high-performance AI solution for the education sector, accelerating the integration of large model technology into education.
Low-cost, High-performance, Promoting Educational Equity
The 'Zi Yue 3 Math Model' has demonstrated significant advantages in solving domestic math problems. Through large-scale reinforcement learning and innovative algorithms, its training cost was only $26,000, with inference performance about 15 times that of DeepSeek R1.
Test results show that on datasets such as CK12-math (Internal), GAOKAO-Bench (Math), MathBench (K12), and MATH500, the lightweight 14B 'Zi Yue 3 Math Model' scored higher than general large models like DeepSeek-R1. In the GAOKAO-Bench (Math) framework, which is based on college entrance exam math questions, the model scored as high as 98.5 points.
Besides its excellent performance, the 'Zi Yue 3 Math Model' also has significant cost advantages. It can run on consumer-grade GPUs, greatly reducing the AI application barriers in educational scenarios.
In fact, long-standing issues in primary and secondary education such as resource inequality, lack of personalized tutoring resources, and high costs of explaining complex questions can be alleviated by educational large models. However, building and deploying high-performance large models is very costly, making it difficult for low-income regions to afford.
To address this, the 'Zi Yue 3 Math Model' has achieved a breakthrough in deployment performance, with service costs as low as $0.15 per million tokens, far surpassing the cost-effectiveness of most current general large models. This enables more educational institutions and developers to deploy professional-grade math AI applications at low cost.
NetEase Youdao's Chief Scientist, Duan Yitao, stated, 'For the industry, the Zi Yue 3 Math Model proves the feasibility of building powerful inference models in specific fields at low cost, providing important reference for the development and application of vertical inference large models.'
Accelerating 'AI + Education' Evolution, Zi Yue Unlocks the Value of Educational Large Models
Since launching the domestic first educational large model 'Zi Yue' in July 2023, NetEase Youdao has accelerated its evolution in the 'AI + Education' field, continuously unlocking the application value of large models in education.
From launching over ten vertical applications within a year, to releasing the first domestic inference model 'Zi Yue-o1' with step-by-step explanations earlier this year, and now to the open source of 'Zi Yue 3 Math Model', Youdao has been iterating and optimizing models based on application scenarios, promoting technological innovation and practical deployment of educational large models.
The release and open sourcing of 'Zi Yue 3 Math Model' marks another major breakthrough in lightweight model inference capabilities, sharing Youdao's long-term accumulated technological achievements in educational large models with more developers and practitioners, fostering a collaborative industry ecosystem.
Duan Yitao emphasized, 'The ultimate goal of AI + Education is not to replace teachers, but to leverage technology as a tool for educational equity. We hope the open source of Zi Yue 3 Math Model will attract more developers to explore its application in education.'
Appendix 1: Demo link, try it out
Appendix 2: Model open source link
https://github.com/netease-youdao/Confucius3-Math
Appendix 3: Paper link
https://github.com/netease-youdao/Confucius3-Math/blob/main/Confucius3-Math.pdf