I was going to say, and I say this a little bit too often, a BERT model can do this even faster, with even less memory...