The residential air conditioners (RACs) are increasing rapidly in urban power systems and have been widely considered as good regulation resources for improving the system flexibility and resiliency. However, in practical power systems, it is difficult to comprehensively acquire millions of RACs’ operating data and buildings’ thermal data, which makes the available regulation capacity of RACs tricky to evaluate. To address this issue, this paper proposes a Gaussian Mixture Model (GMM)-based evaluation method by utilizing partial easily observable data. First, a control framework of large-scale RACs is developed to provide regulation services for the power system. Based on the thermal–electrical models of RACs and buildings, a quantification method of the available regulation capacities is proposed under the premise of guaranteeing all the users’ comfortable indoor temperatures. Considering the practical condition of insufficient data acquisition of large-scale heterogeneous RACs, a GMM-based evaluation method is designed to calculate the probability of semi-info RACs’ expected regulation capacities by sampling a small portion of full-info RACs’ characteristics. Moreover, the Expectation Maximization Algorithm and the Bayesian Information Criterion are employed to optimize the multi-dimensional parameters and the component number of the GMM, which significantly improve the evaluation accuracy with lower complexity. The proposed models and methods are verified in a demonstration project on demand response in China.