To accommodate the effects of crustal deformation in the current national static geodetic datum (Taiwan Geodetic Datum 1997 (TWD97)) in SW Taiwan, 221 campaign-mode global positioning system (GPS) stations from 2002 to 2010 were used in this study to generate a surface horizontal velocity model for establishing a semi-dynamic datum in SW Taiwan. An interpolation method, Kriging, and a tectonic block model, DEFNODE, were used to construct the surface horizontal velocity model. Forty-four continuous GPS stations were used to examine the performance of the semi-dynamic datum through exterior validation. The average values of the residual errors obtained using the Kriging method for the north and east components are ±1.9 and ±2.2 mm/year, respectively, whereas those obtained using the block model are ±2.0 and ±2.9 mm/year, respectively. The distribution of residuals greater than 5 mm/year for both models generally corresponds to a high strain rate area derived using the horizontal velocity field. In addition, these residuals may result from deep-seated landslide and active folding or mud diapir in a mudstone area. Similar exterior checking results obtained using the Kriging interpolation method and block model for SW Taiwan indicate a high station density and a relatively satisfactory station spatial coverage. However, the block model is superior to the Kriging method due to the consideration of characteristics of the geological structure in the block model. In addition, result from traditional coordinate transformation was used to compare with the semi-dynamic datum. The results indicate that a semi-dynamic datum is a feasible solution for maintaining the accuracy of TWD97 at an appropriate level over time in Taiwan.