Introducing Mistral OCR 4

Source

• Source name: Mistral AI
• Source kind: blog (product announcement — industry/marketing)
• Authors / organization: Mistral AI
• Venue: Mistral AI News
• Published: 2026-06-23
• Collected: 2026-06-23
• URL: https://mistral.ai/news/ocr-4/
• PDF:
• DOI:

Note: This is a product announcement from an industry vendor. Claims about benchmark performance are self-reported. Treat as marketing material pending independent evaluation.

Plain-Language Brief

Mistral AI released OCR 4, a document intelligence model that extracts text and provides structural metadata: bounding boxes, block classification (titles, tables, equations, signatures), per-page/per-word confidence scores, and support for 170 languages across 10 language groups. It supports API access and single-container on-premises deployment, and integrates with Mistral's RAG search toolkit.

What the Document Says

• OCR 4 outputs: bounding boxes for text localization, block classification (title/table/equation/signature/etc.), inline confidence scores, multi-format support (PDF, DOC, PPT, OpenDocument)
• Multilingual: 170 languages, 10 language groups
• Deployment: API or single-container self-hosted
• Integration target: RAG pipelines, enterprise search, domain-specific retrieval
• Benchmark scores (self-reported): OlmOCRBench 85.20 (top overall), OmniDocBench 93.07, Internal Crawl Multilingual 0.98
• Human preference evaluations: average win rate of 72% for OCR 4 vs. comparators

Key Concepts Mentioned

• Document intelligence
• Bounding boxes and block classification
• Inline confidence scores
• RAG ingestion pipeline
• Self-hosted single-container deployment
• Multilingual OCR
• OlmOCRBench, OmniDocBench benchmarks

Claims Made by the Document

1. OCR 4 achieves top overall score on OlmOCRBench (85.20)
2. Annotators prefer OCR 4 in 72% of human preference evaluations
3. Supports 170 languages and 10 language groups
4. Model leads across all eight language groups tested

Evidence Used

• OlmOCRBench score (external benchmark)
• OmniDocBench score (external benchmark)
• Internal multilingual benchmark
• Human preference evaluations (not independently published)

Methods / Approach

Product release announcement. Benchmark evaluation described but methodology not fully detailed in the post. Self-hosted deployment tested via single-container setup.

Limitations Stated by the Authors

• Benchmark scoring contains artifacts: ground-truth annotation errors, mathematical notation format mismatches, equation segmentation variations, multi-column reading order assumptions, block-type attribution issues
• Model not intended for medical diagnosis, legal judgment, high-stakes financial decisions, or safety-critical systems

Limitations Noted by the Agent

Agent interpretation — not from the source.

• Benchmark scores are self-reported; independent replication not available at time of collection
• Human preference evaluation methodology (annotator count, task diversity, independence) not disclosed
• "Internal Crawl Multilingual" benchmark is a proprietary internal metric — not independently verifiable
• This is a vendor announcement; claims should be treated as marketing until third-party evaluation is available

Open Questions

• How does OCR 4 compare to commercial alternatives (e.g., AWS Textract, Google Document AI) on identical benchmarks?
• What is the latency/throughput profile for the self-hosted deployment?
• How does block classification accuracy vary across document types?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• Relevant to questions about RAG pipeline components and document ingestion
• The confidence-score feature is relevant to retrieval quality discussions
• [Interpretation] Self-hosted single-container deployment lowers the barrier for enterprise on-premises RAG pipelines

Raw Abstract / Excerpt

"Today, we're releasing Mistral OCR 4, featuring bounding boxes, block classification, and inline confidence scores alongside extracted text. The model supports 170 languages across 10 language groups, runs in a single container for fully self-hosted deployments, and serves as an ingestion component for enterprise search, RAG, and domain-specific retrieval pipelines."