In the paper the aggregation algorithm for Wiener system modelling is proposed. The method uses exponential weights to combine (aggregate) models from a given collection of D-lambda linear dynamic and D-m nonlinear static models of the genuine system blocks (that can be a priori hypotheses, other system identification outcomes, etc.). The resulting model has a noteworthy property: it is almost as accurate as the best hypothetical Wiener model built from a given collection. This is caused by the fact that the error introduced by the aggregation routine decreases (in the mean squared sense) with growing number of measurements N as fast as CqN(-1)ln (D-m D-lambda), where C and q are known constants. Simulations confirm theoretical findings and demonstrate applicability of the algorithm in engineering problems for small and moderate measurement data-sets.