Distilling can be done by fast follower closed models too, so this argument against open models doesn't hold up.