Unless they produce the same token output on the face of it, it looks like they're trying to cover for 4.7 not having good model perf?
Unless they produce the same token output on the face of it, it looks like they're trying to cover for 4.7 not having good model perf?