There's some new models that can run locally, but I have no idea what the performance of them is like. I was wondering if that's what was used, given the sensitivity of the data