The model quality assessment problem consists in the a priori estimation of the overall and per-residue accuracy of protein structure predictions. Over the past years, a number of methods have been developed to address this issue and CASP established a prediction category to evaluate their performance in 2006. In 2008 the experiment was repeated and its results are reported here. Participants were invited to infer the correctness of the protein models submitted by the registered automatic servers. Estimates could apply to both whole models and individual amino acids. Groups involved in the tertiary structure prediction categories were also asked to assign local error estimates to each predicted residue in their own models and their results are also discussed here. The correlation between the predicted and observed correctness measures was the basis of the assessment of the results. We observe that consensus-based methods still perform significantly better than those accepting single models, similarly to what was concluded in the previous edition of the experiment.