Mistral releases OCR 4, that structures documents beyond text

Mistral releases OCR 4, a document recognition model that structures text, tables, and signatures with bounding boxes and confidence scores for RAG.

Mistral releases OCR 4, the new version of its document recognition and understanding model. Where previous generations converted a page into clean text and tables, OCR 4 returns a structured representation: each block is localized by a bounding box, classified by type (title, table, equation, or signature), and accompanied by confidence scores, per page and per word. Downstream systems thus know not only what the document says, but also where each element is located and with what reliability it was read.

The company intends this structure for RAG and enterprise search, source-anchored citations, redactions and human verification, as well as agentic workflows like form filling or invoice processing. The model covers 170 languages, with claimed improvements for rare languages, and remains compact enough to run in a single, fully self-hosted container, allowing documents to be kept within one's own environment for residency and compliance reasons.

Regarding pricing, the API costs four dollars per thousand pages, half that for batch processing, with a Document AI layer adding custom structured output. Mistral reports the model as preferred in blind human evaluations against tested systems, with an average win rate of 72 percent and a first-place ranking on the public benchmark OlmOCRBench, while qualifying these scores, which it deems indicative rather than definitive. It is also available on Amazon SageMaker and Microsoft Foundry.