Lumped models treat a catchment as one unit with average inputs; distributed models split it into many cells, each with its own inputs and values.
Hydrological models differ in how they handle space. A lumped model treats the whole catchment as a single unit. Rainfall, evaporation and soil properties are averaged over the area, and the model predicts flow only at the outlet. GR4J and the lumped form of HBV are common examples. A distributed model divides the catchment into many pieces, usually grid cells, each with its own rainfall, soil, land cover and slope. Water is computed in each cell and passed between cells, so the model can give flow, soil moisture or water table depth at any point. MIKE SHE and many grid-based land surface models work this way. Between the two sit semi-distributed models, such as SWAT and HEC-HMS, which split a catchment into sub-catchments or hydrologic response units and treat each one as lumped. The difference in size can be large. A 500 km² catchment run as one unit with 10 parameters has 10 values to set. The same catchment on a 1 km grid has 500 cells, and if each cell carries its own 10 parameters there are 5,000 values, most of which cannot be measured directly. Lumped models are fast, need little data and often match outlet flow as well as much larger models. They suit flow forecasting at a gauge or running thousands of simulations. Distributed models are needed when the question is about place: where floods start, how land use change in one part of the catchment matters, or where groundwater is falling. They can also take satellite and radar data cell by cell. The cost is data, run time and calibration. In practice most distributed parameters are taken from maps and tables, and only a few scaling factors are calibrated. More spatial detail does not guarantee better predictions at the outlet, and many different parameter sets may fit equally well.