Mistral AI introduces its latest model for optical character recognition to the market. Mistral AI reports on its company blog that the third version of its OCR software offers significant performance gains over previous generations. The update specifically targets challenges like complex tables, handwritten notes, and low-quality scans.
The new model achieves a 74 percent win rate compared to its predecessor when processing forms and dense layouts. Developers can convert images or PDF documents into clean markdown text or structured JSON data. To facilitate testing, the company launched the Document AI Playground. This interface allows users to parse documents through a simple drag-and-drop system.
A key technical feature is the reconstruction of table structures. The system uses HTML tags to preserve column hierarchies and merged cells. This ensures that downstream AI agents understand the visual structure of a report as well as the text itself. Mistral OCR 3 handles cursive handwriting and mixed-content annotations layered over printed forms with high fidelity.
The model remains smaller than most competing solutions and is available at a competitive price. Processing costs two dollars per 1,000 pages. Organizations using the Batch API receive a 50 percent discount. Tim Law, IDC Director of Research, notes that organizations gain a competitive advantage by extracting high-fidelity context from their data. Use cases include the digitization of historical archives, automated invoice processing, and the improvement of enterprise search systems.
