Mistral releases OCR 3
Mistral said the smaller model beat its predecessor on forms, scans, tables and handwriting 74% of the time, at $2 per 1,000 pages.
- Models & capabilities
- Colour
Mistral AI released OCR 3, a document-parsing model it described as much smaller than the competing enterprise and AI-native OCR solutions it compared itself against. The company said the model improved handwriting recognition (including cursive and mixed annotations), form detection in dense layouts such as invoices and receipts, robustness to scan artefacts like compression, skew and low resolution, and reconstruction of complex tables with merged cells and column hierarchies, outputting markdown with HTML-based tables.
Mistral reported a 74% overall win rate against its own predecessor, OCR 2, across those four categories using fuzzy-matching evaluation against ground-truth data — a self-reported comparison, not an independent benchmark. The model is priced at $2 per 1,000 pages, $3 per 1,000 pages when structured annotations are used, and roughly $1 per 1,000 pages through the discounted batch API. It is available via API, through a new Document AI Playground in Mistral AI Studio, and as a self-hosted option for organisations with data-privacy requirements.