Probably the second, in the vast majority of PCs local inference should be painfully slow and resource intensive.