This paper focuses on the dynamic and statistical–dynamical downscaling techniques for estimating the precipitation at the stations in the Heihe River basin of Northwest China based on local observations at 14 sites and the outputs from RIEMS2.0 (Regional Climate Model) with a grid resolution of 3×3 km. The precipitation estimated further using MLR (multiple linear regression) and BMA (Bayesian Model Average) with different factor combinations is tested on the assessment indices as RMSE (Root Mean Square Error), correlation coefficient, variance percent, and “negative precipitation bias” with observation. Results show that the precipitation produced by the dynamic model has the largest RMSE, the most significant coherence, and a considerably larger variance than the observation by a factor of 1.5–2. Except for correlation coefficients, the statistical–dynamical downscaling models are optimal, and the statistical models are between statistical–dynamical models and dynamic model. The test shows that the correlation coefficient of the statistical downscaling models constructed with 700-hPa geopotential height field, meridional wind, and specific humidity is lower, and the RMSE is larger. The statistic indices were improved when the precipitation factor was introduced into the statistically downscaling models. The correlation coefficient and variance percentage of MLR models are considerably higher than BMA models, the RMSE of the two types of models is close in value, but the bias of negative precipitation of the former is significantly higher than that of the latter. The negative precipitation produced by the statistically downscaling models appears mainly in the cold season or dry and arid lands, such as the lower reaches of the river, of which the “negative precipitation” frequency decreases if the model precipitation is added as a factor in the downscaling models. Moreover, the statistical assessment of the monthly precipitation estimated from the downscaling models reveals that the four indices would evolve with season, in which the errors of dynamical downscaling are also the largest among the downscaling models, and their relative errors are smaller in summer and larger in winter, particularly in lower reaches of the river. This implies that precipitation downscaling in the dry land or dry season is still difficult for climate study. These results show a significant bias in dynamic downscaling, even for the high-resolution regional climate model. Therefore, the regional model must be combined with statistic downscaling to form a statistical–dynamical model for decreasing the precipitation uncertainties estimated in the river basin.