How would this be any different from running one of the lower parameter models?
> at minimal loss of accuracy
Suggesting that there is a lot of redundancy in the weights.
As some other commentator stated, there's currently a lot of low hanging fruit in optimizing NN.