I'm pretty sure fast speculative decoding models need to have the same objective (NTP, MDM, NCM) as the larger model they try to accelerate.