Mistral AI Launches OCR 4, Turning Unstructured Documents Into Layered Data

Mistral AI has officially launched its latest innovation, Mistral OCR 4, a sophisticated document intelligence model aimed at revolutionizing how businesses interact with unstructured data. While the model itself was released on June 23, 2026, news outlets are reporting on its capabilities and implications today, June 30, 2026. This new offering significantly advances beyond traditional Optical Character Recognition (OCR) by not merely extracting raw text but by returning fully structured representations of entire documents. Mistral OCR 4 provides granular details such as paragraph-level bounding boxes, allowing for precise location of text within a document. It also includes typed-block labels, categorizing content into elements like titles, tables, equations, and even signatures, alongside confidence scores for each extraction. This rich, layered data is crucial for applications requiring deep document understanding and precise information retrieval. A key advantage of OCR 4 is its native ingestion capability for common unstructured enterprise formats, including PDF, DOC, PPT, and OpenDocument (ODT) files. This eliminates the need for intermediate conversions, streamlining workflows and enhancing efficiency. Furthermore, the model boasts extensive language support, covering 170 languages across 10 distinct groups, demonstrating strong performance even in rare, specialized, and low-resource languages. In rigorous evaluations, Mistral OCR 4 has shown impressive results. Blind evaluations involving over 600 real-world documents revealed that independent annotators preferred OCR 4 over competing systems, achieving an average win rate of 72%. It also secured top positions on automated benchmarks, scoring 85.20 on the public OlmOCRBench, 93.07 on OmniDocBench, and 0.98 on Mistral's internal Crawl Multilingual evaluation. The model integrates seamlessly as an ingestion layer for Mistral's open-source Search Toolkit, providing clean, citation-ready text units directly to RAG and agentic workflows. For regulated enterprise clients, OCR 4 offers flexible deployment options, including a single, fully self-hosted container, ensuring data control and security. Mistral emphasizes that OCR 4 is a document-understanding tool, not an autonomous decision-maker, thus setting clear boundaries for its application, particularly in high-stakes automations like medical diagnoses. The model is immediately accessible via the Mistral API, with various pricing tiers available for different usage patterns.