With the great effects of global warming since the 20th century, extreme climatic events such as storm, drought, flood and heat wave occur more frequently in most regions of the world with longer duration and wider ranges of influence. Given the substantial economic and social consequences, the study of such events has become an important issue in climatic and the related sciences. Most basins in humid and semi-humid south and east of China were concerned for probability modeling analysis of precipitation extremes. While, for the inland river basins, there was a limited presence of such studies partly due to the limited data availability and a relatively low mean annual precipitation. The objective of this study was to carry out probability modeling of precipitation extremes in Heihe River basin, the second largest inland river basin in China, by using the peak over threshold (POT) method and Generalized Pareto Distribution (GPD), in which the selection of threshold and inherent assumptions for POT series were elaborated in details. Percentile method, Hill plot and the average annual occurrence number were used to select the threshold in GPD. L moment was used to estimate the shape and scale parameters. The inherent stationary assumption including no significant trend and no change point in POT series was verified by using Mann-Kendall test and Pettitt test. For comparison, another widely used probability distribution Gamma was employed as well. Results showed that, POT series derived from the aggregated methods were reasonable for the study area and approximately stationary based on the two tests. GPD could give satisfactory fits to the POT series for each station according to the Kolmogorov-Smirnov test, the theoretical and empirical cumulative probability curves and the return levels in different return periods. By investigating the theoretical cumulative frequency over 90%, it is found, in the 1960s, more extreme daily precipitation events were observed; in the 1970s and 1980s, the extreme daily precipitation events became less, while they became common in and after the 1990s. From the perspective of spatial scale, it happened more frequently in the lower reach than in the middle and upper reaches of the Heihe River basin according to the past 51 years (1960-2010) records.