Forth is arguably the best language in terms of expressive power/language complexity, though that's largely due to the denominator.
It's a great choice for tightly constrained devices, though far fewer devices have such constraints now. The mental overhead of having to track stack-effects for each function makes it difficult to scale to large systems and maintain productivity, but if you're trying to eke every iota of power out of a chip with a tiny amount of ROM, it's hard to beat a direct-threaded (or token-threaded...) Forth.