Présentation de Mistral OCR 3

Mistral's new OCR 3 model extracts text and images from documents with remarkable fidelity, excelling at challenging content like handwriting, complex forms, and low-quality scans. It significantly outperforms its predecessor, boasting a 74% higher win rate on benchmarks for forms, scanned documents, and tables. The model intelligently reconstructs document structure by outputting clean Markdown and HTML tables, preserving the original layout for downstream systems. Accessible via API or a user-friendly Document AI Playground, it also offers a self-hosted option for organizations with strict data privacy requirements.