because theano use g++ to compile the model in c code so it runs a bit faster. However, if you are deploying the model, maybe there is a way run the model once where g++ is available, and then copy the generated c code somewhere.
I have never actually done it, so I am not sure if it works or not.