There's a bunch of reasons, but they really come down to a focus on memory efficiency and knowledge of how certain things perform on Python (using hasattr instead of isinstance where applicable, using __slots__, only processing things when it's clear they need to be processed). And then for extra measure: Cython compilation. You can see a competing frameworks benchmarks showing the result of all this work here: http://pycnic.nullism.com/

That's really cool, thanks.