MathVC is a novel virtual classroom platform powered by large language models (LLMs) designed to simulate middle-school students collaborating on mathematical modeling (MM) tasks. The platform aims to address the challenge of uneven distribution of teachers and educational resources, which often limits students' opportunities for practicing MM skills. MathVC features multiple LLM-simulated student characters with specified traits and properties, allowing human students to engage in collaborative problem-solving without the need for teacher monitoring. To enhance the realism and effectiveness of the simulations, the authors propose three key innovations: integrating MM domain knowledge into the simulation, defining symbolic schemas as the foundation for character simulation, and designing a meta planner to guide the conversational procedure. Through experiments and ablation studies, the authors demonstrate the effectiveness of their approach, showing that MathVC can achieve better alignment with authentic student behavior and conversational procedures compared to vanilla simulations. The results suggest that MathVC has significant potential for enhancing mathematics education, particularly in resource-limited settings, and can be deployed as part of the middle-school curriculum to improve students' MM skills.MathVC is a novel virtual classroom platform powered by large language models (LLMs) designed to simulate middle-school students collaborating on mathematical modeling (MM) tasks. The platform aims to address the challenge of uneven distribution of teachers and educational resources, which often limits students' opportunities for practicing MM skills. MathVC features multiple LLM-simulated student characters with specified traits and properties, allowing human students to engage in collaborative problem-solving without the need for teacher monitoring. To enhance the realism and effectiveness of the simulations, the authors propose three key innovations: integrating MM domain knowledge into the simulation, defining symbolic schemas as the foundation for character simulation, and designing a meta planner to guide the conversational procedure. Through experiments and ablation studies, the authors demonstrate the effectiveness of their approach, showing that MathVC can achieve better alignment with authentic student behavior and conversational procedures compared to vanilla simulations. The results suggest that MathVC has significant potential for enhancing mathematics education, particularly in resource-limited settings, and can be deployed as part of the middle-school curriculum to improve students' MM skills.