The pretraining is shared between models, inference is generally such more expensive.