Introducing the OCR model 'Mistral OCR 3,' which can recognize handwritten cursive characters and convert them into text

AI developer Mistral AI has announced the release of Mistral OCR 3, an AI model that can recognize scanned documents and handwritten characters and convert them into text. Mistral OCR 3 boasts accuracy that surpasses that of enterprise document processing solutions and AI-based OCR solutions, and can even recognize hard-to-read cursive writing.

Introducing Mistral OCR 3 | Mistral AI

https://mistral.ai/news/mistral-ocr-3

Mistral OCR 3 is a significant upgrade over its predecessor, Mistral OCR 2, and is now capable of accurately reading handwritten content, including annotations on cursive and printed text. It also boasts improved detection capabilities for dense layouts and forms, making it suitable for reading a variety of contract documents. Furthermore, it is more robust against character compression and distortion caused by low-quality scans, low resolution, and background noise.

It supports Markdown output with HTML-based table structure, allowing downstream systems to understand the document's structure as well as its content. The video below demonstrates Mistral OCR 3 accurately reading letters and numbers while preserving the table structure of the scanned data.

Mistral 3 OCR Demo in AI Studio - YouTube


The graph below compares the performance of Mistral OCR 3 (orange) against competitors such as Microsoft Azure, Google Document AI, and DeepSeek OCR, as measured by various benchmarks. Mistral OCR 3 outperforms competitors in Chinese, East Asian languages, Eastern European languages, English, and Western European languages.

Below is a comparison of the performance of various formats, including input forms, handwritten characters, invoices, complex tables, and scanned data of historical documents, with Amazon Textract added. It can be seen that Mistral OCR 3's performance is superior to that of competing products.

Below is a comparison of the performance of each language with the previous model, Mistral OCR 2. Mistral OCR 3 outperforms the previous model in all languages, achieving a 74% win rate overall.

Mistral OCR 3 can be accessed through the API or Document AI Playground interface within Mistral AI Studio . Smaller than many competing solutions, it's available at an industry-leading price of $2 per 1,000 pages, or $1 per 1,000 pages with the Batch-API discount.