We propose a new Mallows'Cp type criterion for quantile regression (QR) which, unlike AIC or BIC, does not require parametric assumptions on the population and by construction is robust against misspecification. We show that our new Mallows type criterion for QR (QCp) is not only an asymptotically unbiased estimate of the average weighted squared error on the model average fit, but also asymptotically optimal in the sense of achieving the lowest possible weighted squared error in a class of discrete model sets. We also demonstrate that these asymptotic properties of the QCp estimator hold in finite samples with a simulation experiment.