OmniDocBench

Mistral releases OCR 4 with bounding boxes, block classification, and SOTA benchmark scores

Mistral AI released OCR 4, a document intelligence model supporting 170 languages across 10 language groups, with structured output including bounding boxes, typed-block classification, and inline confidence scores. The model claims top scores on OlmOCRBench (85.20) and wins 72% of head-to-head human preference evaluations against competing OCR and document-AI systems. It is deployable in a single container for self-hosted, data-sovereign environments and is priced at $4 per 1,000 pages via API. OCR 4 is integrated with Mistral's open-source Search Toolkit as an ingestion component for RAG and enterprise search pipelines.