The duration distribution of 408 GRBs with measured both duration T 90 and redshift z is examined. Mixtures of a number of distributions (standard normal, skewnormal, sinh-arcsinh, and alpha-skew-normal) are fitted to the observed and intrinsic durations using the maximum loglikelihood method. The best fit is chosen via the Akaike information criterion. The aim of this work is to assess the presence of the presumed intermediate GRB class, and to provide a phenomenological model more appropriate than the common mixture of standard Gaussians. While log T obs 90 are well described by a truly trimodal fit, after moving to the rest frame the statistically most significant fit is unimodal. To trace the source of this discrepancy, 334 GRBs observed only by Swift/BAT are examined in the same way. In the observer frame, this results in a number of statistically plausible descriptions, being uni-and bimodal, and with the number of components ranging from one to three. After moving to the rest frame, no unambiguous conclusions may be put forward. It is concluded that the size of the sample is not big enough to infer reliably GRB properties based on a univariate statistical reasoning only.