By taking sounding and auto station data as “standard data”, both subjective and objective verifications are carried out on the initials and predictions of numerical prediction model GRAPES and WRF to reveal the two models’ initial value quality, dynamical frame performance and rain parameterization in Southwest China to some extent. The results show that the initial geopotential height, wind speed and direction of GRAPES are somewhat better than WRF. In the predicting stage, however, the height and temperature RMS (Root Mean Square error) of GRAPES increase much faster than WRF. Besides, GRAPES predicts the 500 hPa height in Southwest China to be lower systematically, while WRF predicts higher height in Southwest China in a relatively big probability. The verification on the sorted weather processes exhibits that GRAPES is superior to WRF in vortex and shear initial analysis. But in predicting stage, WRF performances remarkably better than GRAPES as WRF predicts vortex, shear and trough in a higher correct rate. The analysis of sorted rainfall displays that for both models, rainfall prediction in vortex process is more difficult than the trough process, especially to GRAPES, but for WRF, rainfall prediction in shear process is the most difficult. This conclusion is accorded with two models’ capability in sorted weather process, so the rainfall parameterization effects in the two models are similar.