← Models

Mistral OCR 4

Mistral AI · Mistral OCR

Mistral's document model for structured extraction, bounding boxes, and confidence-aware pipelines.

Type
multimodal
Context
N/A
Max Output
N/A
Status
current
API Access
Yes
License
proprietary
ocr document-ai multimodal structured-output bounding-boxes confidence-scores multilingual rag
Released June 2026 · Updated July 10, 2026

Overview

Freshness note: Model capabilities, access, and pricing can change quickly. This profile is a point-in-time snapshot last verified on July 10, 2026.

Mistral OCR 4 is Mistral AI’s current document-intelligence model for turning complex files into structured, location-aware content. Released on June 23, 2026, it adds bounding boxes, typed block labels, and inline confidence scores alongside extracted text and Markdown. Mistral positions it as the ingestion layer for enterprise search, retrieval-augmented generation, and agent workflows rather than as a general-purpose decision model.

The API uses the mistral-ocr-latest alias, while Mistral’s model card lists a versioned OCR 4 revision. Production systems should pin a tested version when reproducibility matters instead of silently following the moving alias.

Capabilities

OCR 4 accepts common enterprise document formats including PDF, DOC, PPT, and OpenDocument. It returns page structure as well as content: bounding boxes locate text, block classifications distinguish elements such as titles, tables, equations, and signatures, and confidence scores support review or fallback rules at page and word level.

Mistral reports support for 170 languages across ten language groups in the OCR 4 launch material. That coverage is useful for multilingual archives, but teams should evaluate their own scripts, scans, layouts, and low-resource languages rather than assuming the headline count implies equal accuracy everywhere.

The structured output integrates directly with Mistral Search Toolkit, giving retrieval pipelines citation-ready inputs before chunking, indexing, and evaluation.

Technical Details

OCR 4 is a document-processing model, so token-native context and output limits are not meaningful here. Its contextWindow: 0 and maxOutput: 0 values mean N/A rather than literal token ceilings for a page-priced service.

Pure OCR mode and Document AI use the same underlying OCR engine. The base route returns extracted content, boxes, block types, confidence values, and structured Markdown; Document AI adds schema-driven extraction, custom prompts, and application-level structuring. Mistral classifies OCR 4 as a proprietary Premier model. Enterprise self-managed deployment is available, but there are no public model weights, so it is not an open-source release.

Pricing & Access

Published pricing at this snapshot is:

  • OCR API: $4 per 1,000 pages
  • Batch API: $2 per 1,000 pages after the documented 50% batch discount
  • Document AI: $5 per 1,000 pages

Access is available through Mistral’s API and Mistral Studio, with current provider availability through Amazon SageMaker and Microsoft Foundry. Mistral lists Snowflake support as coming soon, not generally available. Enterprise customers can discuss a self-hosted single-container deployment.

Best Use Cases

Choose OCR 4 for multilingual document ingestion, complex-layout parsing, archive digitization, RAG preparation, redaction support, and extraction workflows that need coordinates or confidence-aware review. It is especially relevant when downstream systems must cite or highlight the exact source region.

Do not treat OCR output as verified truth. For invoices, contracts, clinical material, financial records, or other high-impact documents, route low-confidence fields and consequential decisions to a human or a domain-specific validation step.

Comparisons

  • Mistral OCR 3: The direct predecessor; OCR 4 adds richer segmentation, block types, and confidence-aware output.
  • Google Document AI and Azure AI Document Intelligence: Broader cloud document platforms with mature ecosystem integrations; OCR 4 is a focused alternative with a clear API and optional Mistral-native search pipeline.
  • General multimodal models: Useful for flexible document questions, but a dedicated parser is usually easier to evaluate and operate for repeatable ingestion.

Mistral reports strong human-preference and benchmark results, but those remain provider-run evaluations. Test representative documents, failure classes, latency, and total pipeline cost before choosing a production parser.