ABSTRACTPhysical habitat models are used to relate the availability of suitable habitat to river flows by describing the combined availability of depths, velocities and substrates deemed suitable for a species. Field data are often used to calibrate hydraulic models and therefore calculate availability of suitable physical habitat over a range of flows at a site. This paper describes and tests a parameterize-then-cluster-then-predict strategy that extends physical habitat models across New Zealand. Mixed-effects models were used to represent habitat–flow relationships at 274 sites for each of 17 sets of habitat suitability criteria. Parameter values for each set of habitat suitability criteria and for each formulation were clustered into groups. Cluster membership was related to catchment characteristics. Cross-validated generalized habitat predictions were tested for their ability to predict: (1) availability of suitable physical habitat across a range of flows; and (2) loss of suitable physical habitat between two reference flows. Of the 17 sets of habitat suitability criteria, 12 had predicted habitat–flow patterns that exceeded a set of a priori defined performance criteria, providing evidence that predictions were able to distinguish between-site patterns in relationships between availability of suitable physical habitat and flow at unvisited sites.