Compiling it is straightforward under linux, unless you are missing the boost program_options library, pthreads, or zlib. We've also tested the present code on a mac and found it works well.
The code is in open development on github. You can always grab the very latest version with
git clone git://github.com/JohnLangford/vowpal_wabbit.gitVersion 4.1 is mostly bugfixes on version 4.0. Version 4.0 has many changes, the primary one being experimental support for cluster parallel learning. The hash function has also switched to murmurhash2, and the input language has changed slightly.
Version 3.10 is the previous release under a BSD (revised) license. There have been many changes since the previous release. Most of these involve new features, or rewriting core pieces to perform better. One change that effects users is that the learning rate is now rescaled by the --initial_t parameter, because that makes more sense. Another user-visible change is that we dropped specialist EG for now, as the marginal occasional performance gains didn't appear to justify the computational and code complexity cost.
Version 2.3 was the previous release.