We consider the asymptotics of the Bayesian procedure and more specifically, the asymptotic behaviour of posterior distributions in mis-specified models. Given a prior and an i.i.d P_0-distributed sample (where it is not assumed that P_0 lies in the support of the prior), we show that the posterior concentrates its mass in neighbourhoods of the subset of models where the Kullback-Leibler divergence with respect to P_0 is minimal, under an entropy condition and a prior-mass condition that determine the rate of convergence. The method is applied to a model of Gaussian location mixtures.