Mistral OCR 4

2026-06-23 · Publisher : Mistral AI · Domain : Artificial Intelligence

Mistral AI launches Mistral OCR 4, its document recognition model: 170 languages, coordinates and block types, confidence scores, with a self-hosting option for enterprises.

What changes

  • Bounding boxes for each element of the page and classification of blocks (titles, tables, equations, signatures).
  • Built-in confidence scores and structured output in Markdown.
  • 170 languages grouped into 10 families; PDF, DOC, PPT and OpenDocument formats.
  • Self-hosted deployment in a single container for confidential documents.

### Specifications - Successor to Mistral OCR 3 (December 2025). License: Premier (proprietary, API or self-hosting under contract). - Availability: Mistral Studio (API), Amazon SageMaker, Microsoft Foundry; Snowflake Parse Document announced as coming soon. - API identifier: mistral-ocr-4 (alias mistral-ocr-latest, later redirected to OCR 4.1 on July 16). - Context and cutoff: not applicable or not disclosed. ### Pricing | Item | Price | |---|---| | API | 4 USD per 1,000 pages | | API with Batch (50% discount) | 2 USD per 1,000 pages | | Document AI (annotations) | 5 USD per 1,000 pages | ### Benchmarks | Benchmark | Score | |---|---| | OlmOCRBench | 85.20 | | OmniDocBench | 93.07 | | Human preference versus competitors | 72% wins on average | Scores published by the vendor. ### Key takeaways Block localization and confidence scores make integration easier into RAG and document processing pipelines with traceability, including on-premises.

Links

Tags : LLM, Mistral, OCR, Document AI, Multilingue