Mistral OCR 4: A Complete Guide to Structured Document Understanding

170 languages across 10 language groups. Mistral reports particular accuracy gains on rarer and less-resourced scripts, including Japanese, Hindi, and Greek .

Single-container self-hosting. The model can be deployed entirely on-premises in one container, a key differentiator for regulated industries that cannot send documents to external APIs .

Multimodal input and structured output. OCR 4 accepts PDFs and images (Office documents via conversion) and outputs structured Markdown and JSON, designed for integration with RAG and agentic pipelines .

Benchmark Performance

Mistral also reports strong scores on its internal Crawl Multilingual benchmark, though the raw numbers were not published in the sources reviewed .

Pricing Options

Pricing is page-based, not token-based, which is unusual among Mistral's other models and reflects the document-batch use case.

Strategic Context

OCR 4 marks a deliberate shift from "text extraction" to "document understanding." It is positioned as a fundamental layer for enterprise search, RAG pipelines, and agentic workflows where preserving layout and structure (tables, equations, signatures) is critical . It directly targets Google's Document AI, Azure Document Intelligence, and open-source OCR pipelines by combining structured output at commodity pricing with a self-hostable container option—a rare combination among major OCR APIs .

The single-container design is a key differentiator for finance, healthcare, and legal sectors that require data sovereignty . Mistral has also made OCR 4 the default document-understanding model across millions of users on Le Chat .