Motivation: Multiple alignment of highly divergent sequences is a challenging problem for which available programs tend to show poor performance. Generally, this is due to a scoring function that does not describe biological reality accurately enough or a heuristic that cannot explore solution space efficiently enough. In this respect, we present a new program, Align-m, that uses a non-progressive local approach to guide a global alignment. Results: Two large test sets were used that represent the entire SCOP classification and cover sequence similarities between 0 and 50% identity. Performance was compared with the publicly available algorithms ClustalW, T-Coffee and DiAlign. In general, Align-m has comparable or slightly higher accuracy in terms of correctly aligned residues, especially for distantly related sequences. Importantly, it aligns much fewer residues incorrectly, with average differences of over 15% compared with some of the other algorithms. Availability: Align-m and the test sets are available at http://bioinformatics.vub.ac.be