Imagine Models are small enough to run on consumer hardware so inference doesn't matter. We do need p2p for training and distribution though