Mistral OCR 4 Release Notes: SOTA Document Intelligence

Mistral OCR 4 is a state-of-the-art document intelligence model designed to extract and structure content from complex documents. It moves beyond simple text extraction by providing a structured representation of documents, including bounding boxes for localization, typed-block classification (such as tables, equations, and signatures), and inline confidence scores.

Key Technical Capabilities

Mistral OCR 4 transforms pages into structured data to support advanced downstream AI workflows. The model provides three primary structural primitives:

  • Bounding Boxes: Localizes text to enable in-context highlighting and reliable data pipelines.
  • Block Classification: Identifies the role of content (e.g., titles, tables, equations), which is critical for semantic chunking in Retrieval-Augmented Generation (RAG) and agentic workflows like invoice processing.
  • Confidence Scores: Generates per-page and per-word scores to facilitate human-in-the-loop verification and redactions.

The model supports 170 languages across 10 language groups and accepts common enterprise formats including PDF, DOC, PPT, and OpenDocument. For organizations with strict data residency or sovereignty requirements, OCR 4 is compact enough to be deployed in a single container for fully self-hosted environments.

Performance and Benchmarks

In head-to-head human preference evaluations across 600+ documents in 12+ languages, independent annotators preferred OCR 4 over all leading OCR and document-AI systems tested, with an average win rate of 72%.

Benchmark Scores

  • OlmOCRBench: 85.20 (Top overall score among tested models).
  • OmniDocBench: 93.07.
  • Crawl Multilingual Evaluation: .98 (Leading across all eight language groups).

Mistral notes that automated benchmarks often contain "ground-truth errors" or fail to account for equivalent LaTeX math notations and multi-column reading orders, which can penalize correct output. Consequently, the company treats these aggregate scores as directional rather than definitive.

Integration and Pricing

OCR 4 is integrated with the Mistral Search Toolkit, an open-source framework where the model's structured output serves as citation-ready input for RAG and enterprise search pipelines.

Deployment Options

Developers can access the model through the Mistral Studio API, Amazon SageMaker, and Microsoft Foundry, with Snowflake Parse Document support coming soon.

Pricing Structure

Service Price (per 1,000 pages)
OCR 4 API $4
OCR 4 Batch API $2
Document AI $5

Pure Extraction vs. Document AI

Users can choose between two modes via the same API endpoint:

  1. Pure Extraction: Returns raw extracted content, bounding boxes, and confidence scores for custom downstream logic.
  2. Document AI: Layers additional capabilities on top of the OCR output, such as returning structured JSON based on a user-defined schema or using a custom prompt to summarize the document via the mistral-small-2603 model.

Community Insights and Counterpoints

While Mistral reports high performance, community feedback from Hacker News provides a more nuanced view of the model's real-world application:

  • Language and Accuracy: Some users reported issues with specific languages, with one user noting that a specific style of Malayalam handwriting was misidentified as Kannada.
  • Formatting Issues: A user reported that the model incorrectly converted US double quotation marks to UK-style single quotation marks, ignoring the source document.
  • **Cost Comparisons: Some developers argue the pricing is high compared to alternatives, citing Google Vision OCR at $1.50 per 1,000 pages.
  • Benchmark Skepticism: Some users expressed skepticism regarding internal benchmarks and the use of truncated y-axes in reporting charts.
  • Positive Use Cases: Other users reported high satisfaction when processing 55-year-old degraded paper files and complex magazine layouts, noting it outperformed legacy tools like Abbyy Finereader.

"Mistral OCR is roughly 4x faster per page than our incumbent provider, an impressive result for the high-volume docketing workflows where speed is critical to managing our customers' IP timelines." — Ivan Mihailov, AI engineer, Anaqua

Sources

Related

  • Dispatch
    Mistral OCR 4 release notes / what's new

    Mistral AI has released Mistral OCR 4, a high-performance document parsing model supporting 170 languages and providing structured outputs including bounding boxes, block classification, and confidence scores.

  • Dispatch
    Mistral OCR Release Notes

    Mistral AI has launched Mistral OCR, a high-accuracy Optical Character Recognition API designed to extract interleaved text, images, tables, and equations from complex documents for use in RAG systems.

  • Dispatch
    Mistral OCR 3 Release Notes

    Mistral AI has released Mistral OCR 3, a high-fidelity document parsing model that achieves a 74% win rate over its predecessor on complex documents and handwriting, priced at $2 per 1,000 pages.

  • Dispatch
    Mistral OCR 4.1 launch – pricing, features, and community reaction

    Mistral AI released OCR 4.1 on July 16 2026, offering paragraph‑level bounding boxes, structural labels, and a €3.5 / 1000‑page price, sparking mixed reactions about speed, cost, and accuracy on HN.

  • Dispatch
    Mistral Large 2 release notes / what's new

    Mistral AI has released Mistral Large 2, a 123B parameter model designed for high throughput single-node inference with competitive performance in coding, reasoning, and multilingual capabilities.