Mistral AI launches OCR 4 document intelligence model on Microsoft Foundry and AWS

Mistral AI released Mistral OCR 4, an advanced document-intelligence model that extracts structured data from unstructured documents across 170 languages. The model provides paragraph-level bounding boxes, typed-block labels, and confidence scores — features specifically designed to improve retrieval-augmented generation (RAG) pipelines, where accurate document parsing directly determines answer quality.

The structured-output approach matters for enterprise RAG: bounding boxes and typed blocks let downstream systems preserve document layout and semantics rather than flattening everything to raw text, while confidence scores enable selective human review of low-certainty extractions. The 170-language coverage targets global enterprises with multilingual document estates.

Mistral OCR 4 is immediately available via the Mistral API, Mistral Studio, Amazon SageMaker, and Microsoft Foundry, priced at $4 per 1,000 pages for standard API access. The multi-cloud distribution — landing on both AWS and Microsoft's platforms simultaneously — reflects Mistral's strategy of being available wherever enterprises already run, rather than forcing platform lock-in.

Document intelligence is a competitive and commercially important niche, contested by cloud-native OCR services, specialized vendors, and general-purpose vision models. Mistral's bet is that a purpose-built, RAG-optimized OCR model with structured outputs beats general vision models on accuracy and beats legacy OCR on language coverage and layout understanding. Separately, a Nature study on clinical drug-report generation referenced Mistral models among the architectures used, reflecting growing adoption of open Mistral models in healthcare documentation. Watch for independent accuracy comparisons against incumbent OCR services and whether the $4/1K-page pricing proves competitive at enterprise scale.