Mistral OCR 4 brings multilingual structured document extraction and improved performance

Mistral has released Mistral OCR 4, advancing document understanding with support for bounding boxes, block classification, and inline confidence scores. Each extracted content block is now localized, classified by type, and accompanied by per-page and per-word confidence metrics, alongside the textual output.

The model expands accessibility by supporting 170 languages across 10 language groups, including those that are rare or low-resource, addressing a gap in many existing solutions. Building on these enhancements, Mistral OCR 4 accepts common enterprise document formats: PDF, DOC, PPT, and OpenDocument, broadening its suitability for corporate workflows.

For deployment, Mistral OCR 4 runs as a single container and can be fully self-hosted. This design enables organizations to manage cost-sensitive or high-volume operations while maintaining strict data sovereignty by keeping document processing within on-premises infrastructure.

These capabilities allow the model to serve not only as a text extractor, but also as an ingestion component for enterprise search, retrieval-augmented generation (RAG) systems, and domain-specific retrieval workflows. Mistral OCRv4 and the related Document AI solution can be accessed through APIs on Mistral Studio, Amazon SageMaker, and Microsoft Foundry, with Snowflake Parse Document integration expected soon.