The BSP system was develpoed for very early computers, where every thing was to be handled by the CPU. If any polygon would intersect another then those involved polygons are divide into many. The basis of BSP is, no two polygons will intersect each other. And one can see that there is a rise in no of polygons, when two polygons or blocks intersect.

In case of models, their visible polygons, their vertex rotation, transformation, and z-sorting all the things that are now handled by GPU. Mdl models are handled by GPU now so they are way faster. Where as BSP levels are still handled by the CPU and the resulting polygons are then given to GPU for projection, this is done every frame. Because CPU has to decide what polygons are to be drawn in BSP's case, while models are directly handled by GPU, so the later is way faster.