Mistral AI Unveils "OCR 4" Document Intelligence Model — Structured Output and Self-Hosting
On June 23, 2026, Mistral AI unveiled “OCR 4,” a document intelligence model. The model converts unstructured documents into structured data — complete with bounding boxes, typed-block classification, and confidence scores — and supports single-container self-hosting, positioning it for regulated enterprises that cannot route sensitive documents to third-party cloud APIs.
Details
- Structured output: Classifies titles, tables, equations, signatures, and more by block type, returning bounding boxes and per-word confidence scores alongside extracted text
- Multilingual support: Covers 170 languages across 10 language groups
- Performance: Scores 85.20 on OlmOCRBench, with human evaluators preferring it over competing document-AI systems at an average 72% win rate
- Self-hosting: Deployable in a single container, enabling fully on-premises operation for organizations that cannot send data externally
- Pricing: API access costs $4 per 1,000 pages ($2 with a 50% Batch API discount); Document AI is $5 per 1,000 pages
- Availability: Offered via API through Mistral Studio, Amazon SageMaker, and Microsoft Foundry, with Snowflake Parse Document support coming soon
- Enterprise integration: Also integrated into Search Toolkit, the enterprise search framework announced at AI Now Summit 2026, enabling an end-to-end pipeline from document ingestion to citation-ready RAG
How to try it
- A hands-on guide is available via the “Getting Started with OCR 4 Cookbook”
- Interested users can register for the production webinar held on July 7, 2026, or contact the sales team directly
- See Mistral’s official site (mistral.ai/news/ocr-4/) for further details