It's a bit deceptive to state inference speeds without mentioning the additional things you're doing to achieve them