SUMMARYThe second part of this review on streamflow génération analyses how the knowledge available from field studies (see Part 1) has been used since the 1960s or could be used to improve catchment modelling. After a présentation of the main mode! types, the various problems encountered during the modelling process are discussed.The large variety of hydrologie models available for event or continuous simulation can be reduced to a few main types according to the ways the functional, spatial and temporal aspects of the catchment behaviour are represented. Lumped "blackbox" models are useful for many engineering problems but can not be used in "extrapolation" and give no information on the internai catchment dynamics. Lumped conceptual models, which consider a catchment as a System of interconnected réservoirs and simulate the main global fluxes, use empirical lumped relationships and parameters that often hâve no great physical meaning and are not measurable. Semi-distributed conceptual models use the same réservoir description, but at the scale of "homogeneous" units derived from a space discrétisation, which allows one to take catchment structure explicitly into account. Physicaliy-based distributed models, which use theoretical équa-tions and measurable parameters, provide a dynamic explanation of catchment behaviour but require too much information and are too complex to be easily used at the catchment scale. Physico-conceptual semi-distributed models try to overcome the limits of the previous types, while keeping their advantages, by simplifying the dynamic approach and discrétisation using new concepts.Physically-based or conceptual models, which describe or explain the water cycle at the catchment scale, are very useful for research, but their use in prac- Correspondance : ambroise@geographie.u-strasbg.fr Partie 1 publiée dans Rev. Sci. Eau 11/4 (1998). Les commentaires seront reçus jusqu'au 30 septembre 1999.