I can shave a couple of ms off by having every jar stored instead of zipped. (The Inflater.inflateBytes takes some times.)
Most time is spend in Class.defineClass1. I don't think we can do much about that except "-noverify".
Another notable is XML parsing. It uses String.charAt a lot, which can be improved by rewriting woodstox to use CharSequence which you back with an implementation that has no offset (thus no checks). That gives you 20-40% char handling performance improvements.