Mistral OCR 4 release notes / what's new

Mistral AI has announced the release of Mistral OCR 4, a specialized document parsing model designed for enterprise-grade document intelligence. The model provides a breakthrough in document extraction by moving beyond simple text conversion to a structured representation of documents, providing bounding boxes, block classification, and inline confidence scores.

Technical Capabilities and Structured Output

OCR 4 transforms documents into structured data. Unlike previous generations that focused on clean text and tables, OCR 4 provides the following metadata for every extracted element:

Bounding Boxes: Localizes text for in-context highlighting and reliable data pipelines.
Block Classification: Identifies the role of content blocks (e.g., titles, tables, equations, signatures).
Confidence Scores: Generates inline scores per-page and per-word to facilitate human-in-the-loop verification and source-grounded citations.

These structural primitives allow downstream systems to understand not only the content of a document, but its spatial arrangement and the role of each element. This is specifically designed to support semantic chunking for RAG, agentic workflows (such as form filling and invoice processing), and consistent ingestion for data connectors.

Performance and Benchmarks

Mistral AI reports that OCR 4 outperforms leading OCR and document-AI systems in both human preference and automated benchmarks.

Human Preference Evaluations

In a head-to-head blind ranking of 600+ documents across 12+ languages, independent annotators preferred OCR 4 over every leading system tested, with win rates averaging 72%.

Automated Benchmarks

OCR 4 achieved the top overall score on OlmOCRBench (85.20) and led the internal Crawl Multilingual evaluation (.98). It also scored 93.07 on OmniDocBench.

Mistral AI notes that these automated scores are directional rather than definitive due to known limitations in benchmark scoring artifacts, such as ground-truth errors in reference annotations, equivalent LaTeX math notation, and multi-column reading order assumptions that can penalize correct output.

Multilingual Support

OCR 4 supports 170 languages across 10 language groups, including specialized and low-resource languages. According to internal evaluations, the model maintains high accuracy where many competing systems degrade.

Deployment and Integration

OCR 4 is a compact model that can be deployed in a single container, allowing enterprise customers to keep document data within their own infrastructure for data residency and compliance.

API and Document AI Options

Developers can integrate the model via a single API endpoint with two primary usage modes:

Pure Extraction Mode: Provides raw extracted content, bounding boxes, block types, and confidence scores. This is the best option for high-volume batch ingestion via the Batch API.
Document AI Capabilities: By adding specific parameters, the OCR output is fed to mistral-small-2603 to generate structured JSON based on a user-defined schema, annotate images with a vision-language model, or summarize content based on a custom prompt.

Pricing

OCR 4 API: $4 per 1,000 pages (with a 50% discount for Batch-API, reducing cost to $2 per 1,000 pages).
Document AI: $5 per 1,000 pages.

Ecosystem and Availability

OCR 4 is integrated with the Mistral Search Toolkit, an open-source composable search framework. It serves as an ingestion component that provides citation-ready inputs for RAG and enterprise search pipelines.

The model is available via Mistral Studio, Amazon SageMaker, Microsoft Foundry, and will soon be available on Snowflake Parse Document. Self-hosting options are are available for enterprise customers.