While the rmsd cutoff for maxsub is theoretically adjustable in Rosetta – so which one you get may be dependent on exactly where you’re getting the number from – the defaults I’m seeing for most of the maxsub versions are 4.0. If you’re talking about the scores output by the score application, that’s what you’ll get for the plain “maxsub” score. “maxsub2.0” simply means using an rmsd cutoff of 2.0 versus the 4.0 default.
Unfortunately, it doesn’t look like there’s an easy way to get a 3.5 version of the maxsub, at least not with the commandline Rosetta. If you can use PyRosetta, you can simply pass the rmsd cutoff to the core.scoring.CA_maxsub() function as the third parameter, after the two poses you’re trying to compare.
Regarding differences from other servers, the differences in cutoff probably plays a large role, but there are also differences you’ll see based on slight variations in alignment protocols. If things are right on a line, small differences in how you do the alignment may push them one way or the other, affecting the maxsub value.