The success of the molecular-replacement method for solving protein structures from experimental diffraction data depends on the availability of a suitable search model. Typically, this is derived from a previously solved structure, sometimes by homology modelling. Very recently, Baker, Read and coworkers have demonstrated a successful molecular-replacement case based on an ab initio model generated by ROSETTA [Qian et al. (2007), Nature (London), 450, 259-264]. In this contribution, a number of additional test cases in which ab initio models generated using modest computational resources give correct molecular-replacement solutions are reported. Unsuccessful cases are also reported for comparison and the factors influencing the success of this route to structure solution are discussed.