Rainfall interception is a critical source of water in arid and semiarid mountain forests. Interception loss (I) in mountain forests determines the ecohydrological function of a watershed, and this is particularly important in arid and semi-arid regions that rely on runoff yield to balance human and ecosystem water needs. Here, we measured gross rainfall (P), throughfall (Tf), and stemflow (Sf) from May 2020 to September 2021, and leaf area index (LAI), canopy structural parameters, and meteorological data within a Picea crassifolia (P. crassifolia) forest in the Qilian Mountains of northwestern China. Throughfall and canopy interception accounted for 56.4 and 35.2% of the 745.5 mm cumulative rainfall input during the two growing seasons, respectively. Stemflow of up to 13.6% of total rainfall events occurred when the amount of P > 30 mm, and canopy storage capacity (S) was 2.0 mm. Based on the parameters derived from a regression equation between Tf vs. P and a trial-and-error calibration scheme, three of the physical models, the Návar, reformulated Gash, and Liu, performed very well both for event-based (NSE > 0.8) and total-based (NSE > 0.92) I; a larger deviation was found during large rainfall events, especially for the reformulated Gash model. These results indicated that the power Návar and the reformulated Liu model are best for event-based and total-based I modeling, respectively, in spruce forests in this semi-arid region.