I found the same. If you use the available executables (ver. 681, as available in SWAT website), the differences are really huge (for example check annual average watershed values in "output.std"). The SWAT website based debug version would result in a very high surface runoff (unreliable in my view as I know the watershed response quite well) compared to that from release version. I then complied the source code myself using Visual Fortran Complier, the differences between release and debug versions are rather small. So, I suggest you to use release version of the SWAT website based executable OR, compile source code yourself.