Already trained, but still need some ~200GB GPU mem to run the model.