We have tested the performance of a large set of kinetic energy density functionals of the local density approximation (LDA), the gradient expansion approximation (GEA), and the generalized gradient approximation (GGA) for the calculation of interaction energies within a subsystem approach to density functional theory. Our results have been obtained with a new implementation of interaction energies for frozen-density embedding into the Amsterdam Density Functional program. We present data for a representative sample of 39 intermolecular complexes and 15 transition metal coordination compounds with interaction energies spanning the range from -1 to -783 kcal/mol. This is the first time that kinetic energy functionals have been tested for such strong interaction energies as the ligand-metal bonds in the investigated coordination compounds. We confirm earlier work that GGA functionals offer an improvement over the LDA and are particularly well suited for weak interactions like hydrogen bonds. We do, however, not find a particular reason to prefer any of the GGA functionals over another. Functionals derived from the GEA in general perform worse for all of the weaker interactions and cannot be recommended. An unexpectedly good performance is found for the coordination compounds, in particular with the GEA-derived functionals. However, the presently available kinetic energy functionals cannot be applied in cases in which a density redistribution between the subsystems leads to strongly overlapping subsystem electron densities. © 2009 American Chemical Society.