Mistral launches OCR 4, turning unstructured documents into layered data

Mistral AI released Mistral OCR 4, a document intelligence model that goes well beyond basic text extraction to return fully structured document representations. The system provides paragraph-level bounding boxes, typed-block labels (titles, tables, figures), and per-element confidence scores, and supports 170 languages — positioning it as an enterprise-grade ingestion layer rather than a simple OCR tool.

Crucially, OCR 4 directly processes formats like PDF, DOC, and PPT, making it well-suited for advanced retrieval-augmented generation (RAG) pipelines where clean, structured input determines downstream quality. It is available via API, Mistral Studio, Amazon SageMaker, and Microsoft Foundry, giving enterprises multiple deployment paths across the major clouds.

Document intelligence is a fast-growing enterprise battleground, with competition from Google Document AI, AWS Textract, and specialized startups. Mistral's angle is breadth of language support and typed structural output, which matters for multilingual enterprises and complex layouts. As a European lab, Mistral also benefits from data-sovereignty preferences among EU customers.

The practical value hinges on real-world accuracy on messy documents — scanned forms, dense tables, multi-column layouts — where OCR systems traditionally struggle. Confidence scores are a welcome addition for building reliable pipelines, letting teams route low-confidence extractions to human review. Buyers should benchmark OCR 4 against incumbents on their own document corpus before committing, but the structured-output approach is exactly what RAG builders have been asking for.