Mistral OCR 4 adds structured outputs to raw text extraction. It returns bounding boxes for each text block, classifies blocks as titles, tables, equations, signatures and more, and gives per-word confidence scores. The model handles 170 languages across ten groups, including low-resource ones where rivals struggle. You can deploy it in a single container on your own infrastructure, meeting data-sovereignty and compliance needs while processing batches cost-effectively. In head-to-head human evaluations over 600+ documents in 12+ languages, independent annotators picked OCR 4 over every tested competitor, with average win rates of 72%. On the public OlmOCRBench it scored 85.20, and on OmniDocBench 93.07, though the team notes both benchmarks can undercount correct outputs because of mismatched references, equation formatting differences and multi-column ordering quirks. An internal Crawl Multilingual test gave it a .98 score and showed it leading across eight language groups from English and Western European to Chinese, Southeast Asian and specialized scripts like Bengali and Tamil. OCR 4 plugs into Mistral’s open-source Search Toolkit (public preview) to feed citation-ready, typed blocks into retrieval-and-generation workflows. It accepts PDFs, DOC, PPT and OpenDocument files. Teams can call it via API at $4 per 1,000 pages—or $2 with the 50% batch discount—or use the no-code Document AI interface in Mistral Studio at $5 per 1,000 pages. Independent benchmarks show OCR 4 runs roughly four times faster than some enterprise providers on high-volume workflows, and about eight times cheaper with seventeen-times lower latency compared to leading agentic document parsers on chart-and-figure-dense finance datasets. The release aims to give enterprises a self-hosted, high-throughput OCR engine that supplies not just text but the structure and confidence needed for search, retrieval-augmented generation and automated document handling.