The advantage of stack-based is that instructions need not tell where there arguments live, making them smaller (if you have 16 integer registers, for example, a simple “add register I to register j, store result in register k” instruction needs 12 bits to encode the values of i, j, and k. That’s quite a lot if your instructions are 16 bits)
The disadvantage is that you have to move data to where instructions expect to see it before you can do computations. That makes your code larger and slower.
Last I know needing more instructions to do the same is more of a problem than having smaller instructions solves (https://www.usenix.org/events/vee05/full_papers/p153-yunhe.p...), but that’s over 10 years old.
Being stack-based doesn't help JVM efficiency. Modern JVM's take that bytecode and immediately translate it to some other non-stack-based form and then compile it. JVM bytecode is effectively a serialization format for syntax trees. (Stacks are a nice notation for serializing a tree.)
Reference counting is only used for heap allocated values and types that themselves reference values, like pairs and tables.
https://github.com/basic-gongfu/cixl/blob/master/src/cixl/bo...
I'm sure they will at least inspire some good features in others.
There's a newer one called Kitten that is statically typed and uses term rewriting to allow for more normal syntax when you want. It's pretty cool. Too bad none of these languages will ever take off.