Add smaller NN for faster inference

Goes with lhcb/Allen!1487 (closed) .

These are new weights to achieve faster inference keeping almost identical performance.