Model predictive control, as the most popular intelligent advanced control technology in recent years, is increasingly applied to building air conditioning systems to achieve adaptive and energy-efficient operation of the system, however, model predictive control imposes higher requirements on the computational efforts of the model. In this research, a complex coupled heat transfer model of the increasingly popular radiant floor system is analyzed and a full-order model of the radiant floor is developed. Then, after controllability and observability analysis, a fast high-fidelity reduced-order model that can accurately characterize the dynamic thermal performance of the original model is proposed applying the balanced truncation method. The accuracy of the reduced-order model was verified by comparing the radiant floor surface temperature, return water temperature, zone air temperature, and unheated average internal surface temperature with experimental test results, the TRNSYS model, and the full-order model. The computation time of the reduced-order model for open-loop simulation was reduced by 86% to 98% compared to the TRNSYS model. More importantly, the reduced-order model can effectively reduce the complexity of the model predictive control, resulting in a 38% to 78% saving in computation time compared to the full-order model, significantly improving computational efficiency and enhancing the robustness of the model predictive control.