Mistral OCR

About 90% of the world's organizational data is stored in document form. To effectively leverage this unstructured data, Mistral has unveiled Mistral OCR, which precisely understands every element of a document.

Mistral OCR takes images and PDFs as input and extracts text, media, tables, formulas, and more in order, producing output as interleaved text and images. It particularly excels at deeply understanding documents such as scientific papers containing complex formulas—including LaTeX format—as well as charts and graphs.

This model delivers optimal performance when combined with RAG systems that use slides or complex PDFs as input. Its key features include:

• State-of-the-art understanding of complex document elements
• Native multilingual and multimodal support
• Industry-leading benchmarks and fast processing speed
• Structured output support via the Doc-as-prompt approach
• Self-hosting options for organizations handling sensitive information

Mistral OCR has now been applied as the default document understanding model in Le Chat, and is immediately available via the mistral-ocr-latest API on the developer platform, la Plateforme. Pricing is about 1,000 pages per $1, and using batch inference can make costs roughly 2x more efficient.