MISTRAL DROPS OCR 4 AND SAYS YOU CAN RUN IT WITHOUT SENDING YOUR DOCUMENTS TO ANYONE Mistral AI released OCR 4 this week with a pitch aimed squarely at the regulated industries that have been locked out of most AI document tools: run this entirely inside your own infrastructure. The model ships as a self-contained container that banks, hospitals, law firms, and government agencies can deploy on their own servers without a single page of content ever touching an outside server. The technical capabilities are substantial. OCR 4 covers 170 languages, extracts text with paragraph-level bounding boxes, classifies document blocks by type including titles, tables, equations, and signatures, and adds inline confidence scores throughout. On OlmOCRBench, the standard benchmark for structured document intelligence, it scored 85.20, the highest result in the field. Human annotators preferred it over competing systems at an average 72 percent win rate. Pricing through Mistral’s cloud API runs $4 per 1,000 pages, dropping to $2 with batch processing. The model is also live on Amazon SageMaker and Microsoft Foundry for enterprises that prefer to stay in their existing cloud stack. The real market here is not general OCR. It is the vast pile of enterprise documents that have never been digitized into usable structured data because the only tools good enough to do the job required sending those documents somewhere the company’s legal team would not allow. OCR 4 is built for exactly that customer. Keywords: Mistral OCR 4, enterprise document AI, structured data extraction, self-hosted AI