Built for enterprise teams parsing PDFs and scanned documents at scale, especially regulated industries that cannot send files to cloud-only OCR APIs. In blind testing across 600+ real-world documents, independent reviewers preferred its output over rival OCR systems 72% of the time. It replaces manual document-tagging pipelines and general OCR tools, not vision-language models built for image understanding.
Mistral OCR 4 is Mistral AI's document intelligence model, topping the OlmOCRBench leaderboard at 85.20 as of June 2026. It extracts structured text, pixel-level bounding boxes, typed block classification, and per-word confidence scores from PDFs and scanned documents across 170 languages, and ships as a single container for on-premise deployment in regulated industries.
Provider: Mistral AI · Family: Mistral OCR
More about Mistral AI on HokAI
Input modalities: image, pdf · Output: text
About Mistral OCR 4
Mistral OCR 4 is a document intelligence model developed by Mistral AI and released on June 23, 2026, the fourth generation of Mistral's OCR lineup after OCR 3 (December 2025), OCR 2, and the original Mistral OCR (March 2025). Unlike Mistral's general-purpose language models, it is purpose-built for extracting structured content from PDFs, scanned images, and multi-page documents. OCR 4 adds three structural primitives the earlier generations lacked together: pixel-level bounding boxes that localize every text block, typed block classification across 14 region types (title, table, equation, signature, and more), and per-word confidence scores that flag low-quality extractions before they reach downstream systems. It supports a broad set of languages across 10 language groups.
On public evaluation benchmarks, OCR 4 tops the OlmOCRBench leaderboard at 85.20 and scores 93.07 on OmniDocBench. Independent human evaluators preferred its output over competing OCR and document-AI systems in blind testing across hundreds of real-world documents spanning a dozen or more languages. Rogo, an AI platform for financial services, reports matching leading agentic document parsers' accuracy on financial QA datasets at roughly 8 times lower cost and 17 times lower latency. On raw throughput, the model processes up to 2,000 pages per minute on a single GPU node.
Processing is page-based rather than token-based: each page is analyzed independently and returns a structured object with extracted text in reading order, block-level bounding boxes, block-type labels, and confidence scores. The pages parameter accepts integer arrays or range strings (for example, '0,2-4' selects pages 0, 2, 3, and 4), and the Batch API processes large document sets asynchronously. The Document AI layer, powered by mistral-small-2603, applies a user-supplied JSON schema to produce schema-conformant extraction in a single API call.
OCR 4 accepts PDF, JPEG, PNG, and TIFF inputs. The include_blocks API flag enables the full structural output; without it, the API returns extracted text and markdown only.
OCR 4 is available on three managed platforms: Mistral's la Plateforme API, Amazon SageMaker, and Microsoft Foundry (Azure's model catalog), each authenticated via an API key or cloud IAM. For regulated industries that cannot let documents leave their infrastructure, OCR 4 also ships as a single Docker container for on-premise deployment, a capability neither AWS Textract nor Google Document AI offers.
Mistral positions OCR 4 as a document extraction tool, explicitly out of scope for medical diagnosis, legal judgment, or high-stakes financial decisions without human review, and unsuited to real-time safety-critical systems or non-document inputs like audio or video. Mistral has signed the EU GPAI Code of Practice and holds SOC 2 Type II and ISO 27001 certifications, with GDPR compliance supported via EU-based hosting in Paris and a Data Processing Addendum covering Regulation (EU) 2016/679.
OCR 4 is a substantial upgrade from OCR 3, which won 74% of head-to-head comparisons against OCR 2 on forms, scanned documents, tables, and handwriting but lacked the bounding boxes, block classification, and confidence scores now standard in OCR 4. Bounding boxes were the most-requested missing feature before this release. Mistral's roadmap points toward extending the Document AI layer with figure extraction and table-to-spreadsheet output, positioning OCR 4 as an enterprise document ingestion layer rather than a standalone OCR tool.
Pricing
$4 per 1,000 pages via standard API. $2 per 1,000 pages via Batch API (50% discount, asynchronous). $5 per 1,000 pages via Document AI in Mistral Studio. Self-hosted enterprise pricing available on request from Mistral sales.
Key Features
- Bounding Boxes: Localizes every text block with pixel-level coordinates, enabling citation highlighting, source attribution, and reliable data pipelines that know exactly where each word sits on the page.
- Typed Block Classification: Labels each extracted region by structural role: title, table, equation, signature, list, image, caption, header, footer, code, references, and aside text. Fourteen block types in reading order.
- Per-Word Confidence Scores: Returns inline confidence scores per page and per word so downstream systems can flag low-quality extractions before they corrupt RAG indexes, databases, or compliance records.
- 170-Language Support: Covers 170 languages across 10 language groups, including low-resource scripts that many competing OCR services support incompletely or not at all.
- Self-Hosted Single-Container Deployment: Deploys as one Docker container inside your own infrastructure or private cloud, giving regulated industries full data sovereignty without routing documents through a third party.
- Document AI Add-On: Routes OCR 4 output through mistral-small-2603 to produce schema-conformant JSON from any document type, eliminating the need for custom post-processing prompt engineering.
Pros
- Mistral reports OCR 4 won 72% of blind human-preference comparisons against rival OCR systems across 600+ real-world documents, per its June 2026 release benchmarks.
- Self-hosted single-container deployment lets regulated industries keep documents inside their own infrastructure, something cloud-only services like Google Document AI and AWS Textract cannot match.
- Broad multilingual coverage extends to low-resource and non-Latin scripts that competing OCR services handle inconsistently.
Cons
- Pricing roughly doubled from the previous generation, a meaningful cost increase for teams already running high-volume OCR 3 pipelines.
- Structured output (bounding boxes, block types, confidence scores) requires setting the include_blocks flag explicitly; it is not returned by default, which catches integrators by surprise.
- Self-hosted enterprise deployment requires a commercial sales agreement rather than a self-serve download, adding procurement time for regulated buyers.
Benchmarks
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does Mistral OCR 4 cost in 2026?
Mistral OCR 4 uses per-page pricing: the standard API costs $4 per 1,000 pages, the Batch API drops to $2 per 1,000 pages for asynchronous jobs, and the Document AI add-on (schema-conformant JSON extraction) costs $5 per 1,000 pages through Mistral Studio. Self-hosted enterprise deployment is negotiated directly with Mistral sales and its pricing is not published.
Can you use Mistral OCR 4 without paying?
Mistral OCR 4 requires payment from the first request: there is no free tier on the standard, Batch, or Document AI endpoints. Teams testing at low volume can try open-source tools like Tesseract or Surya before committing to per-page pricing.
What are the best alternatives to Mistral OCR 4?
North Micro Vision Instruct is a smaller 2.4B open-weight model from Cohere Labs built specifically for document, chart, and OCR understanding, a lighter option for teams that don't need Mistral's bounding-box and confidence-score output. LFM2.5-VL-3B runs entirely on-device, useful when documents cannot leave a local machine at all. For general-purpose reasoning over already-extracted text rather than raw document parsing, Mistral Large 3 is the better fit.
Is Mistral OCR 4 better than North Micro Vision Instruct?
For enterprise-scale document pipelines, yes: OCR 4 returns block-level location data, per-region structural labels, and per-word reliability scores that North Micro Vision Instruct does not expose in the same way, and it ships as a self-hosted container for regulated industries. North Micro Vision Instruct is a much smaller 2.4B open-weight model from Cohere Labs, a better fit for lighter workloads or teams that want open weights. Neither vendor publishes head-to-head benchmarks against the other, so the choice comes down to deployment scale.
How do you set up Mistral OCR 4?
Sign up for an API key on Mistral's la Plateforme, AWS SageMaker, or Microsoft Foundry, then send a document with include_blocks set to true to get bounding boxes and block types back. For simple field extraction like invoice totals, pass a JSON schema to the Document AI endpoint instead of writing custom parsing logic. Enterprises needing the self-hosted container should contact Mistral sales early, since SLA and DPA negotiation typically takes two to four weeks.
Top Alternatives
- North Micro Vision Instruct: Choose Mistral OCR 4 for enterprise-scale extraction with bounding boxes and self-hosted deployment; go with North Micro Vision Instruct for a smaller 2.4B open-weight model covering lighter document and chart understanding.
- Mistral Large 3: Pick Mistral OCR 4 for dedicated document extraction with bounding boxes and confidence scores; pick Mistral Large 3 for general-purpose reasoning once text is already extracted.
- LFM2.5-VL-3B: Mistral OCR 4 suits large-scale enterprise pipelines; LFM2.5-VL-3B suits lightweight, fully on-device use where nothing leaves the local machine.