Due to the modeling of CNC machine thermal error has characters of small sample and discrete data, the combined modeling method was presented by integrating grey forecast and least square-support vector machine. According to the experimental data of machine temperature and thermal error, a grey forecast model and a least square-support vector machine were built respectively, and then a combination model was established by using weight coefficients. Taking the increase of the synthetic grey correlation between experimental data and the combined model’s forecast value as the aim, optimization of weight coefficients was done. A modeling test was designed on a viaduct gantry machining center, and the result showed that the optimal weights-based combined modeling was prior to grey forecast, least square-support vector machine and multiple linear regressions on accuracy and generalization. Application of the combined model on the online compensation for CNC machine thermal error can effectively reduce the influence of thermal error on machine’s precision.