This is interesting! My impression was that SML/NJ was the de facto "standard" Standard ML implementation (maybe analogous to SBCL in the Lisp world), and (FWIW) Gemini agrees, describing it as "the oldest and most widely used". So I'm surprised to see someone stick it in the rear behind Poly/ML and MLton.
I don't really know what to make of that, except that I guess there's a surprisingly vibrant SML ecosystem and don't listen to any one person (myself included) about it: try them all and see which SML implementation is right for you :)