Mistral OCR 4

Mistral OCR 4 is a proprietary document extraction model and service developed by Mistral AI. Released on June 23, 2026, it converts documents and page images into Markdown and structured layout data. Its versioned API identifier is mistral-ocr-4-0; the rolling mistral-ocr-latest alias began pointing to this version on release day.^[1][3]

OCR 4 extends Mistral OCR 3 with paragraph-level bounding boxes, structural block labels, and page- or word-level confidence scores. Mistral also added support for 170 languages across 10 language groups and an enterprise self-managed deployment that it says can run in a single container. The service is intended for document ingestion, search, retrieval-augmented generation, form and invoice extraction, and other document-processing pipelines. It is not a general decision system.^[1]

Scroll sideways for more →

Release and access

Mistral introduced OCR 4 as a focused document-understanding model rather than a general-purpose multimodal assistant. It is available through the Mistral Studio API and through Mistral's Document AI interface, which provides a no-code route for document extraction and schema-based annotation. At launch, Mistral also listed Amazon SageMaker and Microsoft Foundry as access routes. Snowflake Parse Document was described as coming soon, not as an already available deployment.^[1]

The official model card classifies OCR 4 as a Premier v4.0 service. It lists OCR, structured annotations, bounding-box extraction, and document question-answering among the supported document workflows. The model card does not offer downloadable weights. Self-managed deployment is instead an enterprise product arranged with Mistral.^[1][2]

Mistral's changelog is the clearest source for version pinning. It records the release of mistral-ocr-4-0 and the simultaneous move of mistral-ocr-latest to that model. Applications that require a fixed behavior should use the versioned identifier rather than assume that the rolling alias will remain on OCR 4 indefinitely.^[3]

Inputs and request limits

The /v1/ocr endpoint accepts a file chunk, a document URL, or an image URL. Mistral's guides demonstrate public PDF URLs, uploaded files reached through signed URLs, base64-encoded PDFs, image URLs, and base64 images. Version-specific launch material names common enterprise formats including PDF, DOC, PPT, and OpenDocument, while the processor guide also documents DOCX, PPTX, PNG, JPEG, and other supported image or document formats.^[1][4][5]

Callers can process an entire document or select pages. The pages field accepts a list of zero-indexed integers or a comma-separated string containing individual pages and ranges, such as 0,2-4. Uploaded documents submitted to the OCR processor must be no larger than 50 MB and no longer than 1,000 pages. Those are OCR processor limits, even though Mistral's generic file-upload service can have different limits.^[3][4][5]

Mistral does not publish a token context window for OCR 4. The documented operational ceiling is therefore the file and page limit, not an inferred context length from another Mistral model. Schema-based document annotations also invoke a downstream language model, which can add separate practical constraints that are not specified as OCR 4 context limits.^[1][2][6]

OCR response and layout structure

A standard response contains a pages array and identifies the model used and usage information. Each page can include its Markdown transcription, extracted images, tables, hyperlinks, header and footer fields, page dimensions, confidence information, and a structural blocks array. Tables may remain inline in the Markdown or be returned separately in Markdown or HTML form. Image and table placeholders in the text can be matched to entries in their respective arrays.^[4][5]

Block extraction is specific to OCR 4 and later models. Each block is returned in reading order and contains a type, top-left and bottom-right coordinates, and extracted content. The documented types are text, title, list, table, image, equation, caption, code, references, aside_text, header, footer, and signature. Table and image blocks can carry identifiers that connect them to the page's separate table or image entries.^[3][4]

The current endpoint reference lists include_blocks as true by default, although the OCR processor guide still describes it as an option that callers enable. Clients that depend on blocks can set the parameter explicitly instead of relying on a default that may change.^[4][5]

Confidence output is controlled separately through confidence_scores_granularity. The page setting returns average and minimum page confidence scores. The word setting adds a confidence value for each word in page text and table entries. Unlike blocks, the API reference says confidence output defaults to none so that the response stays smaller.^[4][5]

This creates a documentation discrepancy. The launch post says every request returns confidence scores, but the endpoint reference says the caller must choose page or word. The endpoint contract is the more specific implementation source. Applications that require confidence values should request them explicitly. A confidence score is an estimate from the model, not proof that a transcription is correct.^[1][5]

Document AI annotations

Document AI adds schema-driven interpretation to the same OCR endpoint. It is a layer over the extracted document, not a different raw transcription format. A document_annotation request asks for full-document information shaped to a user-supplied JSON schema, and an optional prompt can guide the interpretation. A bbox_annotation request applies a schema to detected boxes or images, such as asking for a chart caption or a structured description.^[1][6]

Mistral's annotations guide explains that full-document annotation sends the OCR Markdown and the first eight extracted image boxes to a vision-capable language model. Bounding-box annotation calls a vision-capable model for the boxes after OCR. The OCR 4 release names mistral-small-2603 as the model used for full-document schema output at launch and says an image annotation triggers another vision-language-model call for each image.^[1][6]

The distinction matters when evaluating output. Raw OCR can misread or misorder source content. The downstream annotation model can then make a second error while selecting fields, interpreting a chart, or conforming to a schema. The structured JSON should therefore be checked against the page-level Markdown and bounding boxes in workflows where mistakes carry material consequences.

Evaluation

Mistral reported a human-preference study using more than 600 documents in more than 12 languages. The documents came from third-party vendors and were selected to represent industry use cases. Independent annotators compared OCR 4 blindly against each competing system, one document at a time. Mistral reported that OCR 4 was preferred at an average rate of 72% across the systems shown in its aggregate chart.^[1]

The study used human judgments instead of exact string matching, which reduces penalties for harmless formatting differences. It does not make the result an independent benchmark: Mistral selected the study design, systems, documents, and presentation, then reported the aggregate. The release says OCR 3 was part of the broader evaluation, but the published aggregate human-preference chart does not provide a paired OCR 4 versus OCR 3 rate.^[1]

Mistral also published three automated results:

Scroll sideways for more →

Mistral warns that these aggregate scores are directional. Its audit found incorrect ground truth, equivalent LaTeX expressions scored as mismatches, equation segmentation differences, multi-column reading-order and hyphenation artifacts, and header or footer attribution problems. It also states that every competitor score in its charts is an internal reproduction. No independent reproduction of the OCR 4 results was published with the release.^[1]

The internal Crawl Multilingual evaluation groups results into English, Western Europe, Eastern Europe, Middle Eastern, Chinese, East Asian, Southeast Asian, and specialized languages. These eight benchmark groups should not be confused with the 10 language groups in the broader 170-language support claim.^[1]

Relationship to OCR 3

OCR 4 replaced OCR 3 as the target of the rolling alias, but the older version can still be selected by its pinned identifier, mistral-ocr-2512. The versions also have different prices and evaluation claims.^[1][3][8]

Scroll sideways for more →

OCR 3's announced 74% result was a win rate over OCR 2 on Mistral's internal evaluation of forms, scans, handwriting, and tables. OCR 4's 72% figure is an average human preference over several external systems. The figures use different comparators and methods, so they do not show that OCR 4 is less accurate than OCR 3.^[1][8]

OCR 4 doubled the raw per-page list price and raised the annotation price from $3 to $5 per 1,000 pages. Its Batch API price preserves a 50% discount. These are Mistral's published prices as of July 24, 2026 and may change independently of the pinned model identifier.^[1][2][8]

Architecture and training disclosure

Mistral calls OCR 4 a small, focused model and says it is compact enough for a single-container deployment. It has not disclosed the parameter count, exact neural architecture, training corpus, training compute, training schedule, or token context window in the release, model card, API documentation, or a public technical paper. No public repository or model weights accompanied the launch.^[1][2]

The annotation workflow does not fill that gap. mistral-small-2603 is named as a downstream model that transforms OCR output into schema-shaped content. It should not be treated as OCR 4's backbone, parameter count, or context window.^[1][6]

Mistral says self-managed deployment is available to enterprise customers and can keep documents inside the customer's environment. The company did not publish a general container image, minimum hardware configuration, or guaranteed throughput for that deployment. Claims about a specific GPU requirement or pages-per-second rate would therefore require a deployment-specific agreement or benchmark.^[1]

Privacy, limitations, and intended use

Mistral's privacy documentation says data sent through the hosted API is not used for model training. Its organization controls also include a setting to allow or disable API data use for service improvement and, for eligible organizations, zero data retention. When zero data retention is enabled, Mistral says troubleshooting and analytics data are not stored. Organizations handling sensitive documents should confirm their actual account settings and contract rather than assume that hosted inference and self-managed inference have identical retention properties.^[7]

A self-managed installation can keep source documents within the customer's environment, which may help with residency or sovereignty requirements. It does not remove the need for access controls, audit logging, secure storage, and review of any downstream system that receives OCR or annotation output.^[1]

Mistral explicitly describes OCR 4 as a document-understanding component, not a decision-maker. It is not intended for medical diagnosis, legal advice or judgment, high-stakes financial decisions, safety-critical systems, real-time or latency-sensitive processing, or non-document inputs such as raw audio and video.^[1]

Practical limitations include the 50 MB and 1,000-page request ceilings, possible OCR errors in unusual layouts or degraded scans, uncertain confidence calibration on a user's own domain, and added error paths when Document AI invokes another model. The vendor benchmark results also do not replace testing on an organization's own languages, templates, handwriting, tables, and quality thresholds.^[1][4][5]

References

1. Mistral AI, "Introducing OCR 4," June 23, 2026. https://mistral.ai/news/ocr-4/ ↩
2. Mistral AI, "OCR 4 Model Card," June 23, 2026. https://docs.mistral.ai/models/model-cards/ocr-4-0 ↩
3. Mistral AI, "Changelog," accessed July 24, 2026. https://docs.mistral.ai/resources/changelogs ↩
4. Mistral AI, "Document AI - OCR Processor," accessed July 24, 2026. https://docs.mistral.ai/studio-api/document-processing/basic_ocr ↩
5. Mistral AI, "OCR API Endpoint," accessed July 24, 2026. https://docs.mistral.ai/api/endpoint/ocr ↩
6. Mistral AI, "Document Annotations," accessed July 24, 2026. https://docs.mistral.ai/studio-api/document-processing/annotations ↩
7. Mistral AI, "Privacy and Data Controls," accessed July 24, 2026. https://docs.mistral.ai/admin/monitor-comply/privacy-data-controls ↩
8. Mistral AI, "Introducing Mistral OCR 3," December 17, 2025. https://mistral.ai/news/mistral-ocr-3/ ↩