Mistral OCR 4
by Mistral AI
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.
Mistral OCR 4 is a document-understanding model from Mistral AI that extracts structured content (bounding boxes, typed-block labels, and per-word confidence scores) from PDFs and office documents across 170 languages, and can run fully self-hosted in a single container. It is built for enterprise teams building RAG, agentic, and enterprise-search pipelines that need citation-ready, verifiable document extraction.
Released on June 23, 2026, Mistral OCR 4 is a document-parsing and understanding model that goes beyond plain text extraction by returning bounding boxes that localize elements, typed-block classification (titles, tables, equations, signatures, and more), inline confidence scores per page and per word, and structured markdown output. It supports 170 languages across 10 language groups and accepts PDF, DOC, PPT, and OpenDocument formats. Mistral reports it tops OlmOCRBench with an 85.20 score and 93.07 on OmniDocBench, and that independent annotators preferred it over every leading OCR and document-AI system tested with an average 72% win rate. OCR 4 is available via the Mistral API and Document AI (with structured outputs for custom or no-code pipelines) and through Amazon SageMaker and Microsoft Foundry, and it can be deployed in a single container for fully self-hosted, private processing that meets data-residency requirements. API pricing is $4 per 1,000 pages ($2 per 1,000 with the 50% Batch API discount), and Document AI is $5 per 1,000 pages. Target use cases include retrieval-augmented generation, agentic workflows such as invoice processing and compliance, contract digitization, and enterprise search with source attribution.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based, Contact for pricing
- Target Market
- CTOs, Enterprise Developers, Data Scientists, ML Engineers, AI Product Teams
- Founded
- 2023
- Headquarters
- Paris, France
Key Features
- ✓Structure-aware extraction
Returns bounding boxes, typed-block labels (titles, tables, equations, signatures), and structured markdown rather than plain text.
- ✓Per-word confidence scores
Provides inline confidence scores per page and per word to support human verification and auditable pipelines.
- ✓170-language support
Handles 170 languages across 10 language groups and PDF, DOC, PPT, and OpenDocument formats.
- ✓Single-container self-hosting
Runs in a single container for fully self-hosted, private deployment to meet data-residency and privacy requirements.
- ✓Benchmark-leading accuracy
Reports a top OlmOCRBench score of 85.20, 93.07 on OmniDocBench, and a 72% average human-preference win rate over competitors.
Capabilities
Use Cases
- •RAG document ingestion
Convert documents into citation-ready structured text for retrieval-augmented generation and enterprise search.
- •Agentic document workflows
Power invoice processing, form filling, and compliance checks with confidence-scored extraction.
- •Private, on-prem extraction
Deploy in a single container to process sensitive documents without leaving the enterprise.
Ideal For
Best For
- ✓Citation-ready document extraction for RAG pipelines
- ✓Self-hosted, private document AI for regulated data
- ✓Invoice, contract, and compliance document processing at scale
Integrations
Deployment
Market Analysis
Pros
- ✓Structure-aware, citation-ready output built for RAG and agents
- ✓Self-hosting option for regulated and private data
- ✓Strong reported benchmark and human-preference results
Cons
- ✗Benchmark and win-rate figures are largely vendor-reported
- ✗Focused on extraction rather than end-to-end workflow
Pricing
API
From $4 per 1,000 pages
- ✓Structured OCR output
- ✓50% Batch API discount ($2 per 1,000 pages)
Document AI
$5 per 1,000 pages
- ✓Structured outputs for custom and no-code pipelines
Self-hosted
Contact for pricing
- ✓Single-container private deployment for enterprise customers
API is $4 per 1,000 pages ($2 with the 50% Batch API discount); Document AI is $5 per 1,000 pages; self-hosted enterprise deployment is available on request.
Sources
This page was written from 2 sources, 1 on domains other than mistral.ai.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
NVIDIA Nemotron 3.5 Lightning
Open 30B mixture-of-experts model tuned for the high-volume execution layer of long-running agents
River AI
Token-metered LoRA fine-tuning and reinforcement learning on open-weight models you keep
Scaled Cognition
APT, a large action model trained to take actions instead of predicting text
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.