I still have yet to find a "Small" model that can use function calls consistently enough to not be frustrating. That is the most noticeable difference I consistently see between even older "SOTA" models and the best performing "SMALL" models (<70b).