Mistral introduced OCR 4 with locations for text regions, block classifications, and confidence information alongside extracted content. The announcement also covers multilingual processing and deployment options for document search and agent workflows.

Context

Coordinates can show a reader where an answer came from, while confidence estimates help prioritize uncertain regions for review. That is more inspectable than a fluent summary detached from its source. Complex tables and reading order still need attention, because a correct word in the wrong structural position can change the meaning.

Sources & authors

  1. Mistral OCR 4 : SOTA OCR for Document Intelligence
    Mistral AI · June 23, 2026