Mistral has Introduces Mistral OCR, the world’s best document understanding API that enhances information abstraction and retrieval. Mistral OCR unlike other model is an Optical Character Recognition API that comprehends each element of documents; media, text, tables, equations, with unprecedented accuracy and cognition. Mistral OCR is a powerful tool that processes images and PDFs to extract content, organizing it into a coherent sequence of text and images. This makes it an excellent choice for use with a Retrieval-Augmented Generation (RAG) system, especially when dealing with multimodal documents like slides or complex PDFs.

Key Features and Capabilities.

1. State-of-the-Art Document Understanding: Mistral OCR excels in comprehending complex document elements, including interleaved imagery, mathematical expressions, tables, and advanced layouts like LaTeX formatting. This makes it ideal for processing rich documents such as scientific papers with charts, graphs, equations, and figures.

2. Multilingual and Multimodal: The API is natively multilingual, capable of parsing, understanding, and transcribing thousands of scripts, fonts, and languages. This versatility is crucial for global organizations and hyperlocal businesses dealing with diverse linguistic backgrounds.

3. Top-Tier Benchmarks: Mistral OCR has outperformed leading OCR models in rigorous benchmark tests, demonstrating superior accuracy across various aspects of document analysis, including math, multilingual content, scanned documents, and tables. Mistral OCR was put on head to head check with other models like GPT-4o-2024-11-20, Gemini-2.0-Flash-001, Gemini-1.5-Pro-002, Gemini-1.5-Flash-002, and Azure OCR in order to establish a fair comparison all the models were tested on Mistral's internal “text-only” test-set containing various publication papers, and PDFs from the web. Mistral OCR 2503 outperformed other models with an overall score of 94.89.

4. Fastest in Its Category: Being lighter than most models, Mistral OCR processes up to 2000 pages per minute on a single node, ensuring rapid document processing even in high-throughput environments.

5. Doc-as-Prompt, Structured Output: The API introduces the use of documents as prompts, enabling precise instructions and structured outputs like JSON. This allows users to extract specific information and chain outputs into downstream function calls, building powerful agents.

6. Self-Hosting Option: For organizations with stringent data privacy requirements, Mistral OCR offers a self-hosting option, ensuring sensitive information remains secure within their infrastructure.

Mistral OCR Availability

Mistral OCR is now available on the developer platform "la Plateforme," with upcoming access through cloud and inference partners, as well as on-premises options. Pricing starts at $1 per 1000 pages, with batch inference offering double the pages per dollar. You can try Mistral OCR for free on Le Chat to experience its capabilities.