Taking a look at the current AXIOM code, I can guarantee that the memory usage is going to be considerably higher than any of the standard document models for whatever portion of the tree actually gets built. The OMNodeImpl class alone has 4 object references, and int, and a boolean, for a minimum size of 21 bytes. The OMElementImpl class adds another 6 object references and 2 ints, for a mimum size of 53 bytes. OMTextImpl adds 7 object references and 2 booleans to OMNodeImpl, for a minimum size of 52 bytes. These are much larger than the equivalent classes from standard document models (JDOM, for instance, has 6 object references in its Element class, for a minimum size of 24 bytes, and just 2 object references in its Text class, for a minimum size of 8 bytes).
Do you still have the XMLBench code you used for testing the M1 AXIOM? It's probably a good idea to try another test run as we're nearing the 1.0 release.