Mistral OCR 4: Document AI for 170 Languages

Mistral AI has released Mistral OCR 4, a document-reading model the French company says outperforms every competitor it tested in the majority of cases. Announced on June 23, 2026, the model pulls text out of PDFs, Word files, and PowerPoint decks, and according to Mistral's own announcement it supports 170 languages, returns the coordinates of every block it finds, and can run entirely on a customer's own infrastructure. That last point is the one most likely to interest the enterprises Mistral is courting.

What Mistral OCR 4 reads and returns

Optical character recognition has come a long way from simply turning a scanned page into a string of characters. Mistral OCR 4 is built to understand the structure of a document, not just its text. According to The Decoder, the model identifies where each element sits on the page and classifies what it is, distinguishing titles from tables, equations, or signatures. It also attaches confidence scores so a downstream system can judge how much to trust a given extraction.

Mistral's announcement spelled out the output format in more detail. The model returns extracted text in markdown structure, bounding boxes that localize where text appears, block classification for titles, tables, equations, or signatures, among similar elements, and inline confidence scores reported per page and per word. Markdown matters here because it preserves a document's hierarchy in a form that both humans and language models can parse, which makes the output easy to feed into a search index or an AI agent without an extra cleanup step. The supported inputs cover the common enterprise formats: PDF, DOC, PPT, or OpenDocument files.

The bounding-box feature is more than a nicety. Knowing the exact location of every piece of text lets an application highlight the source of an answer, redact sensitive regions, or reconstruct a document's layout faithfully. For workflows where provenance matters, such as legal review or compliance, being able to point at the precise spot on the page where a fact came from is a meaningful capability rather than a checkbox.

Confidence scoring deserves its own mention, because it changes how a system can use the output. An OCR model that returns text with no quality signal forces the application to treat every extraction as equally reliable, which is dangerous when a misread digit in a financial document or a dropped negation in a contract can cause real harm. By attaching a confidence value at the page and word level, Mistral OCR 4 lets a pipeline route low-confidence regions to human review while letting high-confidence text flow through automatically. That kind of selective escalation is what makes large-scale automation tolerable in regulated settings, where a blanket assumption of perfection is not an option.

The 170-language reach

Language coverage is where Mistral OCR 4 makes its broadest claim. The company says the model handles 170 languages spread across 10 language groups, naming English, Western European, Eastern European, Middle Eastern, Chinese, East Asian, and Southeast Asian families, along with a set of scripts that have historically been poorly served by OCR. That list includes Hindi, Japanese, Georgian, Bengali, Armenian, Hebrew, Greek, Gujarati, Tamil, Malayalam, Kannada, or Telugu.

Those Indic and other non-Latin scripts are the hard part of OCR, and naming them specifically is a pointed move. Many systems perform well on English and Western European text and then degrade sharply on scripts with complex ligatures, right-to-left ordering, or dense diacritics. A model that holds up across Tamil, Telugu, or Armenian is solving a genuinely difficult problem, and it opens the door to markets and document sets that English-first tools leave behind. Mistral, as a European company with a global enterprise ambition, has a clear commercial reason to make multilingual strength a headline feature rather than an afterthought.

The benchmark numbers behind Mistral OCR 4

Mistral backed the launch with a mix of human-preference results and standard OCR benchmarks. The headline figure, repeated by The Decoder, is that in independent testing of more than 600 documents, reviewers preferred OCR 4's results 72 percent of the time against competing models. The Decoder also noted that the model beats all tested competitors across both benchmarks the company used, though it observed that the specific competitor names were not disclosed in the materials.

On the published benchmark side, Mistral's announcement reported an OlmOCRBench score of 85.20, which it describes as the top result, and an OmniDocBench score of 93.07. OlmOCRBench and OmniDocBench are recognized evaluations for document understanding, so naming them gives the numbers more weight than a bare percentage would. Still, the usual caution applies. A 72 percent human-preference rate is a real and useful signal, but the lack of named competitors makes it impossible to know exactly what OCR 4 was measured against, and vendor-run preference tests tend to be designed in ways that favor the vendor. The benchmark scores are more checkable, since the evaluations are public, but the cleanest verdict will come from independent runs once practitioners get hands-on access.

"Reviewers preferred OCR 4 results 72 percent of the time" across more than 600 documents, though the competing models were not named.The Decoder

Pricing and how it stacks up

Mistral published a clear price list, which is itself notable in a field where document-AI pricing is often opaque. The Decoder reported standard extraction at $4 per 1,000 pages and a batch mode at $2 per 1,000 pages. Mistral's announcement added a third tier: Document AI with structured JSON output at $5 per 1,000 pages. The batch tier, at half the standard rate, is aimed at the large back-catalog jobs where a company needs to process millions of archived pages and can tolerate latency in exchange for a lower bill.

At fractions of a cent per page, the pricing is squarely in the territory where automating document processing becomes cheaper than paying people to do it, which is the threshold that turns OCR from a feature into infrastructure. The structured-output tier is the interesting one for builders, because it moves beyond raw text extraction into turning a document directly into the JSON shape an application expects, which removes a parsing layer that teams usually have to build and maintain themselves.

Mistral's announcement noted that the Document AI capability, activated through additional parameters, also supports custom JSON schemas, image annotation with structured outputs, and custom prompts that let a caller ask the model to interpret or summarize what it reads. That feature set blurs the line between OCR and a full document-understanding model. Instead of extracting text and then sending it to a separate language model for analysis, a developer can ask OCR 4 to return exactly the fields an application needs, already shaped and labeled. For a team building an invoice processor or a contract analyzer, that consolidation can collapse a multi-step pipeline into a single call, which lowers both engineering cost and the number of places a bug can hide.

Transparent, published pricing is also a competitive statement in its own right. Several large document-AI services price through opaque enterprise contracts or bundle the capability into broader cloud platforms, which makes it hard for a buyer to compare options or forecast spend. By posting a flat per-page rate with a clear batch discount, Mistral makes its offering easy to evaluate against alternatives, and it signals confidence that the cost holds up to scrutiny. For procurement teams that have to justify a vendor choice, a number they can put in a spreadsheet is worth more than a sales call.

Self-hosting and the enterprise pitch

The capability most aligned with Mistral's identity is self-hosted deployment. The company's announcement states that self-hosted deployment is available for enterprise customers requiring data sovereignty, and that the model runs on a single container for cost-efficient, high-throughput batch processing. The Decoder, working from a narrower brief, noted that its source material did not specify whether self-hosting was available, a reminder that the company's own page carries detail the secondary coverage did not.

Data sovereignty is the crux of the enterprise argument. Documents are often the most sensitive material a company holds, including contracts, medical records, financial filings, and identity papers. Many organizations, particularly in Europe and in regulated industries, cannot send that material to a third-party cloud API without crossing legal or contractual lines. A model that can be deployed inside a company's own environment, processing documents that never leave the building, removes that obstacle. It also fits Mistral's broader positioning as the European alternative to American AI providers, with an emphasis on openness and on giving customers control over where their data lives. The single-container design lowers the operational cost of running it in-house, which matters because a self-hosted model is only useful if a team can actually stand it up and keep it fed.

The cloud options round out the deployment story for teams that do not need on-premises hosting. Mistral's announcement lists API access through Mistral Studio, Amazon SageMaker, and Microsoft Foundry, which means a customer can reach the model through the cloud platform it already uses rather than adopting a new vendor relationship from scratch. The presence on both Amazon and Microsoft marketplaces is a practical detail that matters for enterprise procurement, since spending through an existing cloud commitment is often far simpler than onboarding a fresh supplier. Mistral is meeting buyers where they already are, while keeping the self-hosted path open for those who cannot use a shared cloud at all.

Where document AI is heading

OCR 4 arrives at a moment when document understanding has become one of the most practical and least glamorous corners of applied AI. The reason is retrieval-augmented generation and agentic workflows. Before a language model can answer questions about a company's documents, those documents have to be turned into clean, structured text, and the quality of that conversion sets a ceiling on everything downstream. A garbled table or a missed signature block becomes a wrong answer three steps later. That dependency has quietly made high-quality OCR a foundational layer for enterprise AI rather than a standalone utility.

Mistral's announcement points directly at this, framing the model's use cases around document processing for search systems and integration into AI agent workflows. The markdown output, the bounding boxes, and the structured JSON tier all read as design choices made for a world where the consumer of the OCR output is another model, not a person. That is the broader shift worth noting. OCR used to end at producing readable text for a human; now it is increasingly the front door to a pipeline that feeds machines.

The competitive field is crowded. Google, Amazon, and Microsoft all offer document-AI services, and a wave of specialized startups has targeted the same problem. Mistral's bet is that a combination of strong multilingual performance, transparent pricing, and a self-hostable model will carve out the customers that the hyperscalers' cloud-only offerings cannot reach, particularly those bound by data-residency rules. Whether the 72 percent preference rate holds up under independent scrutiny will shape how persuasive that bet turns out to be.

The release also continues a line of work rather than starting one. Mistral's announcement describes OCR 4 as a significant advance over OCR 3, the prior version, though it stops short of itemizing exactly which features changed. The pattern of iterative releases matters because OCR quality improves in the unglamorous details: handling a slightly skewed scan, a low-resolution fax, a table that spans two columns, or a handwritten annotation in a margin. Each generation tends to chip away at the long tail of document formats that break earlier models, and a company that ships regularly accumulates that robustness over time. For buyers, a vendor with a visible release cadence is a safer bet than one that ships a single impressive demo and then goes quiet.

There is a strategic dimension for Mistral as well. The company has built its reputation partly on open and accessible models, and document AI is a domain where that philosophy has a concrete payoff. An organization that can inspect and host the model it relies on for its most sensitive paperwork, keeping full control over it, has a fundamentally different relationship with the vendor than one renting an opaque cloud endpoint. By making OCR 4 self-hostable and openly priced, Mistral is turning its broader positioning into a product feature, betting that control and transparency are worth as much to enterprise buyers as raw accuracy. In a market where the largest providers compete mainly on the strength of their clouds, that is a genuinely different angle of attack, and the coming months of real-world deployment will reveal how many customers it wins.

The bottom line on Mistral OCR 4

Mistral OCR 4 is a confident release in a category that has become strategically important without attracting much attention. The 170-language coverage, the structured and localized output, and the option to run the model on a company's own hardware add up to a tool aimed squarely at serious enterprise document processing rather than casual use. The pricing is clear and aggressive, and the published benchmark scores give the launch more substance than a preference percentage alone would.

The caveats are the familiar ones for any vendor-led launch. The 72 percent figure comes without named competitors, the benchmark results are first-party until others reproduce them, and the multilingual claims, while specific, will need testing on real-world documents in the harder scripts before anyone can treat them as settled. None of that detracts from a clear-eyed read of the release: Mistral has shipped a capable, openly priced, self-hostable document model that fits both its own European positioning and the growing need to feed clean text into AI systems. The next step belongs to the practitioners who will run it against their own documents and report what they find.

What makes this release worth watching is less the individual numbers than the direction they point. Document understanding has quietly become the layer that decides whether enterprise AI systems give right answers or wrong ones, because everything downstream inherits the quality of the text extraction. A model that reads 170 languages well, localizes what it finds, scores its own confidence, and can run on hardware a customer controls is built for exactly that role. If Mistral OCR 4 performs in the field the way its benchmarks suggest, it will not be a flashy product, but it may be a foundational one, and foundational tools tend to matter long after the launch-day headlines fade.