Same model, faster inference