Small area estimators based on area-level random effect models are popular. When the normality assumption fails for random effects, the properties of the estimators deteriorate. In these cases, robust versions of small area predictors are useful. As an alternative to robust empirical best linear unbiased predictors, we propose an extension of M-quantile small-area methods to area-level models. We apply our methodology to estimate the mean equivalized income for local labour systems in Italy via data from the EU-SILC survey. The advantages of the proposed technique are demonstrated in the application and in a simulation exercise.