I used a really good Latin OCR tesseract model, and using fabric I’m piping each of the OCR’d Latin into gpt-oss:20b (running locally on my m4 pro) and yeah it takes a bit but we are 25 pages in and it’s mostly coherent and readable.

It’s translating a Catholic text which has never been translated into English. I’m sure not all will go smoothly but testing between different models they all seem to agree that a literalish translation is happening.