【Objective】 Digital Soil Mapping is receiving more attention and becoming widely used. Its methods mainly include environmental correlation-based models, spatial auto-correlation based models, and a mixture of these two kinds of models. The mixed model is expected to be advantageous over the single models. A generalized linear geostatistical model (GLGM) is a kind of mixed model. Compared with the commonly used mixed model, i.e., regression kriging (RK), GLGM has advantages such as having random effects to account for the non-stationarity of soil variability. However, GLGM is seldomly used due to its major disadvantages, i.e., complicated computation. 【Method】 In this study, within a small hilly area (3.03 km²) in Gaofeng Forest of Nanning, Guangxi, generalized linear mixed model (GLMM) and its combination with ordinary kriging (OK), i.e., GLGM, were used to predict the spatial distribution of soil organic carbon (SOC), pH, clay and cation exchange capacity (CEC). Performances of the two models were then compared with commonly used models, including multivariable linear regression (MLR), geographically weighted regression (GWR), regression forest (RF), OK, RK and generalized additive model (GAM). 【Result】 The results showed that GLMM had higher accuracy in predicting clay, while GLMM and GLGM had medium accuracy in predicting CEC. Further, GLMM and GLGM had lower accuracy in predicting SOC and pH. 【Conclusion】 Based on the adjusted R²of the linear regression model, nugget effect and global Moran’s I, it is concluded that GLMM and GLGM are appropriate when there is a low adjusted R² of linear soil-landscape regression (less than 5%), weak spatial auto-correlation of soil (nugget-to-sill ratio large than 71%), and strong local spatial variability of soil (Moran’s I less than 0.09), e.g., clay in this paper. Otherwise, GLMM and GLGM are not appropriate, e.g., for SOC and pH in this paper. For the high spatial heterogeneity and multi-scale variability of soil, we think that GLMM and GLGM are promising for DSM, although more studies are needed to improve the efficiency of GLMM and GLGM modelling.