I'm using MLton for executables and MoscowML for the REPL, but I wish it was easier than that because you have to come up with some tooling to use both in the same project.
That's what's so exciting about new work on MLton like this: A JIT-ing MLton, for instance, could be an amazing thing. It seems like the community has been coalescing around MLton against all odds, and any improvements to it are a big deal.
SML is nicer syntactically than OCaml, but OCaml brings so much to the table that it's hard to ignore (a simpler compilation model, niceties like named arguments, single development target, good community). Also Paulson's ML For the Working Programmer explains SML's warts well, in a historical context. It goes into great detail of some of the limitations of SML's module type safety and some of the nasty things you need to do to make functors generic and reusable, which really turned me off of the language and explained why Caml and OCaml exist. I still love SML but must concede that it's mostly a teaching language. The Little MLer is wonderful.
Ur/Web also lacks incremental compilation, rendering it unusable for general purpose real world applications.
Otherwise, as a modern ML that has improved upon existing MLs (Hakell included), it would be great to see it, 1ML, or other modern spin on SML take root in the FP community.