Land-surface models (LSMs) are increasingly called upon to represent not only the exchanges of energy, water and momentum across the land–atmosphere interface (their original purpose in climate models), but also how ecosystems and water resources respond to climate, atmospheric environment, land-use and land-use change, and how these responses in turn influence land–atmosphere fluxes of carbon dioxide (CO<sub>2</sub>), trace gases and other species that affect the composition and chemistry of the atmosphere. However, the LSMs embedded in state-of-the-art climate models differ in how they represent fundamental aspects of the hydrological and carbon cycles, resulting in large inter-model differences and sometimes faulty predictions. These "third-generation" LSMs respect the close coupling of the carbon and water cycles through plants, but otherwise tend to be under-constrained, and have not taken full advantage of robust hydrological parameterizations that were independently developed in offline models. Benchmarking, combining multiple sources of atmospheric, biospheric and hydrological data, should be a required component of LSM development, but this field has been relatively poorly supported and intermittently pursued. Moreover, benchmarking alone is not sufficient to ensure that models improve. Increasing complexity may increase realism but decrease reliability and robustness, by increasing the number of poorly known model parameters. In contrast, simplifying the representation of complex processes by stochastic parameterization (the representation of unresolved processes by statistical distributions of values) has been shown to improve model reliability and realism in both atmospheric and land-surface modelling contexts. We provide examples for important processes in hydrology (the generation of runoff and flow routing in heterogeneous catchments) and biology (carbon uptake by species-diverse ecosystems). We propose that the way forward for next-generation complex LSMs will include: (a) representations of biological and hydrological processes based on the implementation of multiple internal constraints; (b) systematic application of benchmarking and data assimilation techniques to optimize parameter values and thereby test the structural adequacy of models; and (c) stochastic parameterization of unresolved variability, applied in both the hydrological and the biological domains.