A specialist with a narrow but deep focus: reading receipts and invoices. Donut skips traditional OCR pipelines entirely, parsing document images end-to-end using a vision-language approach. It's fast and accurate on structured commercial documents, but outside that domain it has little to offer.