Using model-only levels has tons of advantages:

1. They render far faster. With my system having a 3 GHz processor and Radeon 9600 XT video card without updated drivers (drivers from August 12, 2004 or December 8, 2004 - date format conflict), I'm able to get over 100,000 visible polygons and still have above 60 fps. That's without LOD, collision, or other complex things either.
2. The build times are near-instant. Regardless of complexity, it can take but a half second for it to actually build so you can build and run your level all within just 5 seconds, probably not even that long.
3. Flexibility in shape. Why bother restricting yourself to convex surfaces when models support any shape whether simple or complex.
4. Otherwise unlimited level size. Blocks don't compile much beyond 120,000 quants from the origin. Yet, models can go beyond 1,000,000 quants, theoretically up to 2,097,152 quants, allowing for extremely large levels.

These are the biggest of advantages. Tossing in LOD, you could simulate a level of 500,000 visible polygons and get above 60 fps. Collision, shaders, and complex dynamic light arrangements could have a further reduction in the frame rate.