Individual-based forest models (IBMs) are useful to investigate the effect of environment on forest structure and dynamics, but they are often restricted to site-specific applications. To build confidence for spatially distributed simulations, model transferability, i.e. the ability of the same model to provide reliable predictions at contrasting sites, has to be thoroughly tested. We tested the transferability of a spatially explicit forest IBM, TROLL, with a trait-based species parameterization and global gridded climate forcing, by applying it to two sites with sharply contrasting climate and floristic compositions across the tropics, one in South America and one in Southeast Asia. We identified which parameters are most influential for model calibration and assessed the model sensitivity to climatic conditions for a given calibration. TROLL produced realistic predictions of forest structure and dynamics at both sites and this necessitates the recalibration of only three parameters, namely photosynthesis efficiency, crown allometry and mortality rate. All three relate to key processes that constrain model transferability and warrant further model development and data acquisition, with mortality being a particular priority of improvement for the current generation of vegetation models. Varying the climatic conditions at both sites demonstrate similar, and expected, model responses: GPP increased with temperature and irradiance, while stem density and aboveground biomass declined as temperature increased. The climate dependence of productivity and biomass was mediated by plant respiration, carbon allocation and mortality, which has implications both on model development and on forecasting of future carbon dynamics. Our detailed examination of forest IBM transferability unveils key processes that need to improve in genericity before reliable large-scale implementations can be envisioned.