Mistral releases OCR 4: bounding boxes, 170 languages and self-hosted deployment

Mistral AI has released OCR 4, an engine for intelligent processing of enterprise documents. Unlike its predecessors, the model doesn't just return extracted text. It produces a full structured map of the document, with block coordinates, block types and confidence scores for every word.

What changed in version 4

The most requested addition is bounding boxes: every text block now gets precise coordinates on the page. Downstream systems can use them to highlight a cited source directly in the interface, build reliable data pipelines and add human-in-the-loop review for sensitive documents. Until now, most OCR tools produced "flat" text with no link to where it sat on the page.

OCR 4 also classifies each block by type: heading, subheading, paragraph, table, equation or image caption. It adds inline confidence scores at both the page level and the level of each individual word. Together, these features enable fundamentally new use cases: citations tied precisely to their source, automatic redaction of confidential data and operator-controlled verification of results.

RAG pipelines benefit the most. Classified blocks make high-quality retrieval units, and agents can do more than read documents: they can act on them by filling out forms, processing invoices and running compliance checks.

Technical specs and pricing

OCR 4 accepts standard enterprise formats, including PDF, DOC, PPT and OpenDocument, and supports 170 languages across 10 language groups. Mistral specifically highlights quality gains for rare and low-resource languages, where most competing systems degrade noticeably.

Key features:

• Bounding boxes — precise localization of every block on the page
• Block typing — headings, tables, equations, captions, images
• Confidence scores — at the page level and for every word
• 170 languages across 10 language groups, including low-resource ones
• Single-container deployment — the entire model fits in one container

API pricing is $4 per thousand pages. Batch processing through the Batch API gets a 50% discount, bringing the cost to $2 per thousand pages. Document AI in Mistral Studio, a no-code interface, costs $5 per thousand pages.

Single-container self-hosted deployment is available to enterprise customers who need data sovereignty, regulatory compliance and high-throughput batch processing. The model's compact size makes it a fit for both budget-conscious use cases and high-load processing.

Benchmarks and integrations

Independent annotators preferred OCR 4 over every OCR and Document AI system tested, with an average win rate of 72%. On the public OlmOCRBench benchmark, the model scored 85.20, the best result among the tested solutions at the time of publication.

"Downstream systems get access not only to what is written in a

document, but also to where each element is located, what role it plays and how confident the model is about each area of the page," is how Mistral describes the philosophy behind the release.

OCR 4 is integrated into Mistral Search Toolkit, an open framework for enterprise search announced at AI Now Summit. It serves as the ingestion component for RAG pipelines and enterprise search: the model's structured output becomes citation-ready input for retrieval, evaluation and result reranking systems.

What it means

Mistral is turning document recognition from a utility into an infrastructure primitive for enterprise AI systems. Structured output with coordinates, block types and confidence scores is exactly the level of detail agentic systems need to work reliably with real-world documents. Companies building RAG platforms and document intelligence products get a ready-made component that requires no additional post-processing.