2. (Almost) painless concurrency.
3. Symbolic programming right in the language (think of SymPy built in the language).
4. Closeness to Matlab (which makes it easy to translate existing code).
In general, people tend to port to Julia algorithms that were originally written in Matlab or C++ (with the conciseness of first and speed of the second). For example, I made several attempts to translate Matlab/C++ code of active appearance models to NumPy, but succeeded only with Julia [1]. Another example - Mocha.jl [2] - is a port of Caffe deep learning framework (written in C++ with a wrapper in Python) into a pure Julia, which it easy both to use and modify.
So if you just use existing libraries, there's probably no big difference what to use. But if you actively write new libraries, at least give Julia a try.