Mistral launches OCR 4 featuring bounding boxes and typed block classification

Mistral AI introduced OCR 4 on June 23, 2026. The compact model extracts document content while supplying bounding boxes for localization, typed-block classification for elements such as titles, tables, equations and signatures, plus inline confidence scores at both page and word levels.

OCR 4 delivers breakthrough benchmark performance. Independent annotators preferred the new release over all leading OCR and document-AI systems evaluated, producing average win rates of 72 percent. It also posted the highest overall score on OlmOCRBench with 85.20. The company highlighted known limitations in those evaluation methods within its release notes.

Unlike earlier versions centered on producing clean text and tables, this iteration returns a structured document view. Each identified block carries spatial coordinates, a functional category, and reliability metrics. Those additions reportedly support tasks including in-context highlighting, source-grounded citations, redactions, and human-in-the-loop review.

Integration with Mistral Search Toolkit is now available in public preview. The model functions as an ingestion layer inside the open-source, composable search framework that Mistral unveiled at the AI Now Summit. Its structured output feeds directly into the toolkit’s retrieval, evaluation, and RAG workflows.

OCR 4 handles standard enterprise file types such as PDF, DOC, PPT, and OpenDocument formats. It shows measurable improvements on rare and low-resource languages where many rival tools reportedly falter. The model fits inside a single container, enabling fully self-hosted operation that satisfies data residency, sovereignty, and compliance needs while permitting efficient batch processing at scale.

Pricing and access paths target both developers and enterprise teams. Access through the API costs $4 per 1,000 pages, with a 50 percent discount available for Batch-API calls. Developers call the service programmatically; non-technical teams can reach the same engine via the no-code Document AI interface in Mistral Studio. Self-managed deployment remains restricted to enterprise customers.

The release also positions the model for semantic chunking inside RAG systems, for agentic primitives that enable form filling or compliance checks, and for consistent typed feeds into indexing pipelines. Mistral provided additional direction on selecting between the model API and Document AI depending on use case.