This paper proposes a generalized radial basis function (GRBF) network to accurately describe drain to source current nonlinearity for intermodulation distortion (IMD) prediction of MESFETs and HEMTs applications in their saturated region. Trying to analytically reproduce the nonlinearities second and third order Taylor‐series coefficients, responsible for IMD performance in these devices, may result in a quite difficult task. Neural networks were introduced as a robust alternative for microwave modeling, mostly employing the black‐box model type approach of the multilayer perceptron network. The GRBF network we consider is a generalization of the RBF network, which takes advantage of problem dependent information. Allowing different variances for each dimension of input space, the GRBF network makes use of soft nonlinear dependence of the drain to source current derivatives with drain to source voltage for improving accuracy at reduced cost. The network structure and its learning algorithm are presented. Results of its performance are compared to other structures with similar amounts of parameters. Carrier to intermodulation (C/I) predictions validate this approach for precise IMD control versus bias and load in class A amplifiers applications. ©1999 John Wiley & Sons, Inc. Int J RF and Microwave CAE 9: 261–276, 1999.