My quick search reveals the reference implementation (http://stax.codehaus.org), the Woodstox implementation (http://woodstox.codehaus.org), and Sun's SJSXP implementation (https://sjsxp.dev.java.net/).
Please comment on the relative merits of these, and fill me in on any other implementations I should consider.
Woodstox wins every time for me. It's not just performance, either - sjsxp is twitchy and overly pedantic, woodstox just gets on with it.