I think the smaller model is at least 20 times smaller. If you do speculative decoding on a 70B model an 1B model would be appropriate.