Mistral develops advanced OCR for scientific and multilingual documents

Mistral has announced the launch of Mistral OCR, an API designed for optical character recognition in digital documents. The tool is aimed at processing documents containing diverse elements such as text, images, tables, and mathematical equations.

According to data presented by the company, approximately 90% of organizational information is stored in document format. The new API processes both images and PDF files and extracts their content while maintaining the original structure, facilitating integration with retrieval-augmented generation (RAG) systems that work with multimodal documents.

The service has already been implemented as the default model for document understanding in Le Chat, Mistral's conversational platform. The company has published comparative test results where Mistral OCR achieves an overall performance of 94.89% compared to solutions such as Google Document AI (83.42%), Azure OCR (89.52%), and GPT-4o (89.77%).

Technical specifications indicate that the system can process up to 2,000 pages per minute on a single node. Another notable feature is the ability to use entire documents as instructions and generate outputs in structured formats such as JSON.

Developers mention various fields where this technology could be applied, such as digitizing scientific research, preserving historical documents, optimizing customer service, and converting technical and educational literature into formats that can be processed by artificial intelligence systems.

The API is available on Mistral's developer platform, called "la Plateforme." The company also offers free trials through Le Chat and contemplates on-premises installation options for organizations with special data privacy requirements.