I just came across this post about XMI from Steven Kelly over at Metacase. In particular, he quotes some figures about usage of the various XMI versions: he's conducted a web search for XMI files out there on the web. The first thing that struck me was how few there are, and secondly how very few there are (34) that use the latest version (2.0) released in 2003.
Steven also makes the observation that XMI is just an XML document containing sufficient information to describe models. Provide your tool stores its models in XML files and/or provides API access to allow you to create models using the API, it's no big deal to write your own importer in code or using XSLT. And if you want to import an XMI file into a domain specific tool, you'd have to do this in any case, because it is very likely your model will be full of stereotypes and tagged values which will need special interpretation that would not be provided by an off-the-shelf importer.
In another post, Steven talks about XMI[DI] which also supports diagram interchange. Another fact: a 20-class diagram takes nearly 400KB. That was a bit of a surprise to me. But it's his observation that "Standards are great, but I think they work best when they arise from best practice based on good theory, not when some committee tries to create something without ever having done it in practice". This strikes a chord with me: it's what I've called premature standardization.