What is the functional difference between GPT-40 and its mini version?

It's stated that the mini version uses less data or fewer credits, and if it's still effective, then that would be a good thing. I'm just wondering if there is any real numerical or functional information about that.