Mistral OCR 4 (https://console.mistral.ai/build/document-ai/ocr-playground and documentation: https://docs.mistral.ai/models/model-cards/ocr-4-0) extracts and structures the content of documents.

It succeeds Mistral OCR 3 by adding the precise location of each element (bounding boxes with coordinates), their classification by type (title, table, equation, image, text) and confidence scores.

Take into account handwriting.
Accepted formats: PDF, DOC, PPT and OpenDocument.
Support for 170 languages.

The Studio’s “Document AI” interface offers three output tabs: plain text, Markdown, and Visual.
The Visual tab includes translation in any of the 170 languages available.

The download produces a global Markdown file and, per page, a subfolder with a Markdown file, the images extracted as JPEGs, and the detected links (hyperlinks.md).

Accessible for free from “Document AI” in the Mistral Studio with a free account.
10 documents maximum of 50 MB each per session.
Quota displayed in the Studio resets every 2 days, number of pages not specified (130 pages processed in test with no impact on the counter).

Source for the tests: 4 chapters of the dossier “ Artificial intelligence in education – Benchmarks, resources and activities for the classroom” of Réseau Canopé, https://www.reseau-canope.fr/ia-et-education/dossier-ia-generatives-en-education