The most capable model you can run on a single GPU or TPU.