This paper focuses on the dynamic and statistic-dynamic downscaling techniques for estimating the precipitation at the stations in Heihe river basin of Northwest China depending on local observations at 14 sites and the outputs from a regional climate model (RIEMS2.0) with a resolution of 3 km×3km grids. The precipitation estimated further respectively by a multiple regression (MLR) and a Bayesian Model Average(BMA) with different factor combination is tested on the assessment indices as the errors, variance, "negative precipitation bias" and correlation coefficient with observation. Results show that the precipitation produced by the dynamic model is of the biggest errors, the most significant coherence, much large variance than observation by factor of 2 about, while big errors, low correlation coefficient and lower variance than observation are estimated by the statistically downscaling model with the factors as geopotential height, v-wind and specific humidity on 700hPa. When the model precipitation is introduced into the statistically downscaling models, the indices become improved, in which the correlation and variance percentage of MLR"s models are much more higher than BMA"s, so do the errors and "negative precipitation bias". The negative precipitation produced by the statistically downscaling models appears manly in cold season or in dry- and extremely dry land such as lower reaches of the river, of which the negative precipitation frequency become decreased if the model precipitation is added as a factor in the downscaling models. Besides, the statistical assessment of the monthly precipitation estimated from the downscaling models reveals that the four indices would be evolving with season, in which the errors of the dynamical downscaling is also the biggest among the downscaling models, and their relative errors are smaller in summer and bigger in cold season, especially in lower reaches of the river. It implies that the precipitation downscaling in dry land or dry season is still a difficult task to be studied further. These results show that a significant bias exists in the dynamic downscaling even for the regional climate model with high resolution. So, the statistic downscaling has to be combined with the regional model for decreasing the uncertainties of the precipitation estimated in the river basin.