Mistral OCR 4 returns locations and confidence alongside extracted text

Mistral released OCR 4 on 23 June, adding bounding boxes, block classification and confidence scores to extracted document text. The model distinguishes elements such as headings, tables and equations, while locating them on a page. Mistral says it supports 170 languages across ten language groups.

Those outputs help a downstream application show the original evidence behind a citation or route uncertain regions for review. They also make parsing errors easier to investigate than a plain text dump. Access is through the model API or Document AI, with self-managed deployment offered to enterprise customers. The company’s benchmark claims include stated scoring limitations; confidence values and a structured output remain model estimates, so an ingestion pipeline should preserve the source pages as well as the extracted result.