The extreme temperature has more outstanding impact on ecology and water resources in arid regions than the average temperature. Using the downscaled daily temperature data from 21 Coupled Model Inter-comparison Project (CMIP) models of NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP) and the observation data, this paper analyzed the changes in temporal and spatiotemporal variation of temperature extremes, i.e., the maximum temperature (Tmax) and minimum temperature (Tmin), in the Kaidu-Kongqi River basin in Northwest China over the period 2020–2050 based on the evaluation of preferred Multi-Model Ensemble (MME). Results showed that the Partial Least Square ensemble mean participated by Preferred Models (PM-PLS) was better representing the temporal change and spatial distribution of temperature extremes during 1961–2005 and was chosen to project the future change. In 2020–2050, the increasing rate of Tmax (Tmin) under RCP (Representative Concentration Pathway) 8.5 will be 2.0 (1.6) times that under RCP4.5, and that of Tmin will be larger than that of Tmax under each corresponding RCP. Tmin will keep contributing more to global warming than Tmax. The spatial distribution characteristics of Tmax and Tmin under the two RCPs will overall the same; but compared to the baseline period (1986–2005), the increments of Tmax and Tmin in plain area will be larger than those in mountainous area. With the emission concentration increased, however, the response of Tmax in mountainous area will be more sensitive than that in plain area, and that of Tmin will be equivalently sensitive in mountainous area and plain area. The impacts induced by Tmin will be universal and far-reaching. Results of spatiotemporal variation of temperature extremes indicate that large increases in the magnitude of warming in the basin may occur in the future. The projections can provide the scientific basis for water and land plan management and disaster prevention and mitigation in the inland river basin.