**[Mistral AI Launches OCR 4 Document Understanding Model, Supporting Self-Hosting and Custom Schema Output]** According to monitoring by Beating Insights, Mistral AI has released its intelligent document understanding model, Mistral OCR 4. In addition to text extraction, the model outputs bounding boxes, element classification (e.g., headings, tables, formulas, signatures), as well as word-level and page-level confidence scores, compatible with mainstream formats such as PDF, DOC, and PPT. OCR 4 achieved a score of 85.20 on the multi-column and mathematical formula benchmark OlmOCRBench, ranking first in the industry, and scored 93.07 on OmniDocBench. Early enterprise user Rogo's testing revealed that in dense chart-based financial Q&A scenarios, the new model's parsing accuracy rivals that of the cutting-edge Agentic parser, while costing only one-eighth and reducing latency to one-seventeenth. The model supports single-container self-hosting, enabling local protection of enterprise data privacy. Standard extraction pricing is $4 per thousand pages, with batch interface usage reducing the cost to $2 per thousand pages. For $5 per thousand pages in Document AI mode, users can pass in custom JSON Schemas, allowing backend models like mistral-small-2603 to directly output formatted data. The model is now available on Mistral Studio, Amazon SageMaker, and Microsoft Foundry. [Original Link]