I've been wondering about this...I can see a large model getting hit by mutliple lights being worse than many small models being hit by single lights, but is there any other reason why performance would suffer? Couldn't it potentially reduce the number of drawcalls?
This is more in regards to the renderer frontend. After the frontend has found out which visleafs will potentially contribute to the current rendering, it will go through the entities to see whether they are visible or not. To speed up this process, each entity has an axis-aligned bounding box it references. This is a simple cuboid containing the entity. As it only consist out of eight vertices, it is pretty easy (and fast) to determine whether the box is in the players view frustrum or not. If not, the entity isn't either, as it is contained inside the cuboid. If you want to take a look at those boxes type r_showViewEntities 1 into the console.
If you have several smaller models, of which some are not inside the players view frustrum, and therefore don't need rendering, the engine will be able to discard them pretty easely, as most of them will have bounding boxes not inside the view frustrum. If those smaller models are merged into one model however, the engine cannot exclude it and has to go through all the tris the model consists of to find out which of them needs rendering and which not. This is bad.
It is true, though, that you may get a lower amount of drawcalls in return. However. Drawcalls aren't the only thing that can cause performance issues. People are focusing way too much on the renderers backend, but the frontend is important, too. Remember the visportal discussion we had a while back (http://forums.thedar.../?hl=visportal)
Also, when helping Biker with the performance for 'The Elixir', a lot of the performance issues were frontend related. So this isn't an academical problem.