You could always start with a big model then periodically prune low activation parameters as part of fine tuning until the desired size is reached.