In order to optimize the stiffness of car body, Latin design of experiment method is taken to design sample for 300 times. Fitting approximate models of response surfaces with moving least squares methodology are inducted in torsion stiffness optimization, in which the disadvantage of lower precision with least squares methodology is avoided and the relative error percentage is less than 3%. Adaptive response surface optimization methodology is used to optimize approximate models. Optimized plan of changing the thicknesses of front beams and front wheel houses mainly is confirmed. As mass increasing by 6.21%, torsion stiffness increases by 34.85%. The thicknesses of some parts are configured reasonably.