>or some sort of cracked way to pack a ton of parametric knowledge into a Flash Model.

More experts with a lower pertentage of active ones -> more sparsity.