A clear understanding of housing demand structure is the fundamental of making housing policies and stipulating relevant technical standards. In particular, it is critical to clarify the dividing lines of households with different demands, so as to define the target of housing policies. In addition, it is significant to estimate the appropriate housing size demand of different demand groups in order to make practical housing provision plans. In this paper, the above issues are addressed through modeling the housing demand of urban families on a micro scale. First, the influencing factors of urban households' housing demand, including family structure, income, current dwelling conditions, social status, location choice, investment tendencies, etc., are analyzed with their interactions being considered. Based on theoretical analysis, a two-level interaction model is proposed for representing the housing demand structure. Second, the model is used for analyzing the housing demand of Beijing residents, and it is empirically proved to be successful. As a result, eight family groups with significantly differentiated housing demand were identified and the thresholds for splitting the family groups, such as monthly family income of 5000 RMB and current housing status, were clarified. Furthermore, the differentiations of housing size demand across the demand groups are quantitatively clarified. It was found that 100 m2, 110 m2, 140 m2, 170 m2, and 230 m2 were currently thought to be "ideal" for families with 1, 2, 3, 4, 5 or more persons, respectively. These ideal sizes are obviously unrealistic considering the restrictions of land and resources of Beijing, therefore they were adjusted based on the modeling results. The calculation implied that the appropriate housing sizes for 2-person, 3-person, 4-person and 5-or-more-person families are 70-80 m2, 80-95 m2, 90-110 m2, and 90-155 m2, respectively. These results provide many useful implications for housing policy and enrich the methodology of housing demand structure analysis through micro-scale modeling.