added a comment - - edited
I did pull down your repo and ran some tests.
This version of k-nucleotide does run faster than the currently submitted version in alioth.
However, we've got a ways to go to catch the faster Java implementations.
I've added a few comments in your repo.
Of note, the (set! unchecked-math true) directive at the top of files allows us to use inc, dec, etc directly without explicitly specifying unchecked-inc, etc.
http://shootout.alioth.debian.org/u64q/program.php?test=knucleotide&lang=java&id=9 is the fastest java version, but is a bit gnarly.
http://shootout.alioth.debian.org/u64q/program.php?test=knucleotide&lang=java&id=2 is a little slower, but in the same order of magnitude for performance, and significantly more straightforward.
You might check those out for some inspiration to boost the performance of clojure's knucleotide.
Good luck and thanks for the help!