Via ☞ Mikeal Rogers:
Currently (CouchDB 0.11.0) there are three built-in reduce functions. Built-in reduce functions are performed right inside CouchDB implemented in Erlang. In most cases it is very fast because they are way more efficient.
J. Chris Anderson explains why built-in reduce functions are faster:
The deal is that map function (and intermediate reduce function) output is cached in the view index (as it is generated). So in those cases the function overhead is not an issue.