Mistral OCR 4 adds locations, block types, and confidence scores

Mistral released OCR 4 on June 23 with bounding boxes, block classifications, and confidence information alongside extracted document text. The announcement says the model supports 170 languages and offers a self-managed deployment option for enterprise customers. Mistral also distinguishes raw extraction from additional Document AI processing that reshapes extracted content.

Document structure becomes output

Locations and block types provide a different integration surface from plain text. They can help an application connect an extracted field to its position on the page or treat a table differently from a heading. Mistral reports benchmark and human-preference results, while acknowledging problems in reference annotations, reading-order assumptions, and equivalent mathematical notation.

Those caveats matter when deciding what an aggregate score can establish. This brief reports the release, not an independent ranking of OCR systems.

Validate the field the workflow uses

Our analysis: a document pipeline should test the downstream contract rather than only transcription quality. An invoice workflow, for example, needs to know whether a total is associated with the correct currency, supplier, and page region. Correct words in the wrong structural grouping can still produce an incorrect business record.

Confidence scores also need calibration against the organization’s documents before they determine automatic acceptance or review. Sample low-confidence and high-confidence output, including scans with rotation, redaction, and complex tables.

Keep the source document and extraction version linked to every generated field. That makes it possible to inspect a disputed result and rerun processing without losing provenance. Before adopting the new output structure, check that existing consumers tolerate extra blocks and that additional interpretation stages remain separately observable.