Tensor-centric storage layer for AI model hubs — compress checkpoints by exploiting their internal structure.