Mistral Unveils OCR 4

Mistral has released Mistral OCR 4, which goes beyond simple text extraction to fully grasp the structure of documents. This model provides text along with Bounding Box, Block Classification, and Inline Confidence Scores, turning a document's visual structure into data.

OCR 4 offers the following key features:

• Structured output: It distinguishes titles, tables, formulas, signatures, and more, returning the role and positional information of each along with the document content.
• Extensive language support: It supports 170 languages across 10 language groups, showing particularly strong performance in low-resource languages.
• Flexible deployment: As a compact model that can run in a single container, it supports Self-hosted deployment for enterprises where data sovereignty is important.
• RAG optimization: The extracted structural data is optimized for Semantic Chunking for RAG (Retrieval-Augmented Generation) and for agents performing document tasks.

In terms of cost, the API model is $4 per 1,000 pages (or $2 when using the Batch API), and the no-code tool Document AI is offered at $5 per 1,000 pages. It has also been integrated as a core component of Mistral's open-source retrieval framework, Search Toolkit, supporting the construction of enterprise search pipelines.