The gravel pile is one of the methods of tackling soft-foundation. Besides by static load test, its bearing capacity is often calculated by empirical formulae. Because of the difference of methods and theories of these formulae, and the bias of concerning physicomechanical parameters, the effect on bearin capacity of many factors is difficult to be reflected, thus it affects the computation accuracy. In order to overcome the limitation and indeterminateness of single model (empirical formula), a composite model of foundation bearing capacity is given based on adaptive nonlinear neural-fuzzy inference system (ANFIS) in MATLAB. The practice shows the prediction relative error of the composite model is not only less than 3% being superior to all of the submodels, but its accuracy may meet the need of practical projects, thus it provides a new method for calculating foundation bearing capacity.