Mistral OCR 3 Release Notes

Mistral AI has announced the release of Mistral OCR 3, a document parsing model designed to extract text and images from diverse document types with high fidelity. The model significantly improves accuracy in processing forms, scanned documents, complex tables, and handwriting, achieving a 74% overall win rate over Mistral OCR 2.

Technical Capabilities and Document Support

Mistral OCR 3 is engineered to handle a wide variety of organizational and everyday document types, moving beyond the specialization of many existing OCR solutions. Key technical upgrades include:

• Handwriting Recognition: The model accurately interprets cursive text, mixed-content annotations, and handwritten entries layered over printed forms.
• Form Processing: Mistral OCR 3 features improved detection of labels, boxes, and handwritten entries within dense layouts, making it suitable for invoices, receipts, and government documents.
• Scanned Document Robustness: The system is more resilient to low DPI, background noise, skew, distortion, and compression artifacts.
• Complex Table Reconstruction: The model reconstructs table structures including merged cells, column hierarchies, and multi-row blocks. It outputs HTML table tags utilizing colspan and rowspan to preserve the original layout.

Integration and Output Formats

Mistral OCR 3 supports markdown output enriched with HTML-based table reconstruction, allowing downstream systems to maintain an understanding of both content and structure.

Developers can integrate the model via the API using the model identifier mistral-ocr-2512. The model is also available through the Document AI Playground in Mistral AI Studio, providing a drag-and-drop interface to parse PDFs and images into clean text or structured JSON.

Pricing and Deployment Options

Mistral OCR 3 is a smaller model compared to most competitive solutions, which enables a competitive pricing structure:

• Standard API Pricing: $2 per 1,000 pages.
• Batch-API Discount: A 50% discount reduces the cost to $1 per 1,000 pages.

For organizations with strict data privacy and regulatory requirements, Mistral AI offers a self-hosting option to ensure sensitive information remains within the organization's own infrastructure.

Recommended Use Cases

Mistral OCR 3 is designed for high-volume enterprise pipelines and interactive workflows, including:

• Converting documents into markdown for knowledge systems and AI agents.
• Automating the parsing of operational documents, invoices, and forms.
• Digitizing historical or handwritten archives.
• Building end-to-end document understanding pipelines.

According to Tim Law, IDC Director of Research for AI and Automation, "OCR remains foundational for enabling generative AI and agentic AI," noting that high-fidelity extraction provides richer context and a competitive advantage for organizational data.