Mistral OCR 4

Mistral OCR 4, released on 23 June 2026, turns documents into structured data rather than into a wall of text. Where earlier optical character recognition handed back a transcript and left the layout behind, this version returns bounding boxes for every paragraph, labels each block by type — title, table, equation, signature — and attaches confidence scores per page and per word, so a downstream system can tell what it should not trust. The multilingual reach is the widest in the field at 170 languages across ten language groups, and Mistral reports the best published score on OlmOCRBench at 85.20 alongside 93.07 on OmniDocBench. The commercially decisive feature is not accuracy but where the model can run: banks, insurers and hospitals can self-host it inside their own infrastructure, so documents that may not leave the building never do. It is also sold through Mistral's API at four dollars per thousand pages, halved to two in batch mode, and through Amazon SageMaker, Microsoft Foundry and Snowflake. OCR 3 remains available for existing integrations.

▸Videos

No videos yet.

▸Reviews