Some undesirable properties of the standard Takagi–Sugeno (TS) inference method are discussed in relation to analysis of the TS model and to its approximation accuracy. A new inference method based on a smoothing maximum function is proposed. This method guarantees smoothness of the model output to a desired degree, boundedness of the output gradient at each point by the local gradients of the rule consequents and also improves the accuracy of the TS model, as demonstrated on numerical examples.