Mistral AI launches OCR 4 with structured output

French artificial intelligence company Mistral AI has released a new document processing model, OCR 4, which moves beyond simple text extraction to provide a structured map of a document's layout and content. The model, which the company says supports 170 languages, is designed for enterprise use cases where data sovereignty and integration into automated workflows are priorities.

The release marks the fourth generation of Mistral's optical character recognition technology. Unlike previous versions that focused on converting pages into clean text, OCR 4 returns a layered representation of a document. This includes bounding boxes that locate each element on the page, classification of blocks by type-such as title, table, equation, or signature-and per-word and per-page confidence scores. The company stated that bounding boxes were its most-requested feature, as they allow downstream systems to trace extracted information back to its exact source, a critical requirement for compliance and audit trails.

The model accepts common enterprise file formats including PDF, DOC, PPT, and OpenDocument. It is available immediately through the Mistral API, Mistral Studio, Amazon SageMaker,...