Confucius3-Math is an open-source large-scale language model with 14 billion parameters that runs efficiently on a single consumer-grade GPU and achieves state-of-the-art performance on a variety of mathematical inference tasks. It focuses on mathematics learning for Chinese K-12 students and educators, and excels at solving key Chinese K-12 mathematics problems in line with the Chinese National Curriculum at low cost through post-training with large-scale reinforcement learning (RL). In this paper, we share the development process, the problems encountered, and the techniques for solving them, and introduce three technical innovations: target entropy regularization, recent sample recovery, and policy-specific difficulty weighting. These innovations include a new entropy regularization, a new data scheduling policy, and an improved group relative advantage estimator, which significantly improve the stability of RL training, improve data efficiency, and enhance performance. This work demonstrates that it is possible to build powerful inference models in specific fields at low cost. The model and code are open sourced on GitHub.
