Mistral OCR 3
OCR 3 extracts text and embedded images from documents. Mistral explicitly keeps it available for existing integrations even though OCR 4 is newer. Its output combines Markdown with HTML table structure, which can preserve information that a plain text extraction would lose.
Retain the structure of complex tables
The release describes table output with HTML colspan and rowspan attributes. It also targets handwritten entries, forms, skewed scans and low-quality documents. These are vendor-described capabilities, rather than a guarantee for every document.Mistral OCR 3 release and document output details
Separate self-hosting from a public download
Mistral offers a self-hosting route through its team. This is not evidence of a public OCR 3 weight repository. The current API page lists OCR, structured annotations and batching.Mistral OCR 3 release and document output detailsOCR 3 API model page
Compare the stored output before migration
Practical guidance. For an existing parser, compare old and new outputs on the same documents before changing the model. Check the reconstructed cells and the rules that turn them into fields. A newer OCR model can improve text recognition while still changing assumptions in a downstream parser.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | API, Provider product |
| Model identifiers | mistral-ocr-2512 |
| API identifier | mistral-ocr-2512OCR 3 API model page |
| Access | OCR API and Document AI interface; commercial self-hosting by arrangementMistral OCR 3 release and document output details |
Related models
- Mistral OCR 4.1
Mistral AI ยท Mistral OCR
A document-extraction model that returns text with paragraph boxes, structural labels and confidence scores.
Launch date API