Mistral has released OCR 4, a major upgrade to its document-parsing model that goes well beyond raw text extraction. The new model returns structured, annotated output — bounding boxes, typed block classifications, and per-word confidence scores — all deployable in a single container on your own infrastructure. For teams building private knowledge bases, retrieval-augmented generation (RAG) systems, or agentic document workflows, this is the kind of foundational component that changes what's architecturally possible.
The announcement, detailed by Mistral on June 23, 2026, positions OCR 4 not just as a better OCR engine but as a structured ingestion layer for enterprise search and AI agent pipelines. Here's what actually changed, why it matters, and what it means for teams running self-hosted AI.
What's New in OCR 4
Previous Mistral OCR generations focused on converting a page into clean text and tables. OCR 4 shifts the paradigm: every extracted element comes wrapped in metadata.
Bounding boxes localize each text block on the page, enabling downstream systems to know *where* something sits — critical for in-context highlighting in document viewers and for building reliable data pipelines that can reference specific regions of a source file.
Block classification assigns a type to each segment: titles, tables, equations, signatures, and more. This means a downstream system doesn't just receive a wall of text — it receives a structured representation of the document, with each block labeled by its semantic role.
Inline confidence scores are generated per-page and per-word. This is the feature that makes human-in-the-loop verification practical at scale: instead of reviewing every document, teams can flag only the low-confidence regions for human review, saving enormous amounts of labor.
Together, these three additions transform OCR from a "text in, text out" utility into a document-structure API — and that distinction is what makes OCR 4 interesting for the self-hosted AI ecosystem.
Benchmark Performance
Mistral evaluated OCR 4 against leading AI-native OCR models, frontier general-purpose models, enterprise document services, and its own predecessor, OCR 3. Two evaluation approaches stand out:
Human preference testing. Mistral assembled over 600 documents across 12+ languages, sourced from third-party vendors to represent real industry use cases. Independent annotators blindly ranked each competitor's output against OCR 4's. Annotators preferred OCR 4 in the majority of documents across all systems tested, with win rates averaging 72%. Because these are human judgments on realistic documents rather than string comparisons against fixed references, they sidestep much of the annotation and formatting noise that affects automated scores.
OlmOCRBench. OCR 4 achieved the top overall score of 85.20 on this benchmark.
A quote from Rogo, a production user, captures the practical impact: independent testing showed "equivalent accuracy at roughly 8x lower cost and 17x lower latency" compared to leading agentic document parsers on a chart-and-figure-dense financial QA dataset. That kind of cost-latency delta compounds fast at production scale.
170 Languages Across 10 Language Groups
OCR 4 supports 170 languages, with measurable gains on specialized and low-resource languages where competing systems degrade. For organizations operating across multiple geographies — or processing documents in languages like Arabic, Thai, Hindi, or less-common European scripts — this broad coverage eliminates the need to chain multiple OCR providers together.
Multilingual support isn't just about text extraction accuracy. When block classification and confidence scores work correctly across 170 languages, the structured output becomes reliable enough to drive automated pipelines in non-English contexts without constant manual fallback.
Single-Container Self-Hosting
Perhaps the most consequential detail for the self-hosted AI community: OCR 4 is compact enough to deploy in a single container. This keeps all document data within the organization's own infrastructure — no documents leave the network, no third-party API sees sensitive content.
This matters for:
- Data sovereignty and compliance. Regulated industries (finance, healthcare, legal) often cannot send documents to external APIs. A self-hosted OCR model in a single Docker container solves that problem cleanly.
- Cost control. Running on your own hardware means no per-page API fees for internal workloads. The API pricing ($4 per 1,000 pages, or $2 via Batch API) offers a fallback for burst capacity, but steady-state processing can stay on-premises.
- Latency. Local inference eliminates network round-trips. For high-volume batch processing — think thousands of PDFs queued overnight — this adds up quickly.
Self-managed deployment is currently available to enterprise customers.
What This Means for RAG and Agentic Workflows
The real significance of OCR 4 lies in how its structured output feeds downstream AI systems.
Semantic chunking for RAG. Clean, classified blocks become better retrieval units than raw text segments split by character count. When each chunk carries a block type (title, table, paragraph), retrieval systems can weight and rank results more intelligently. Confidence scores let RAG pipelines deprioritize uncertain extractions rather than silently hallucinating from garbled OCR.
Structural primitives for agents. AI agents that process documents — filling forms, processing invoices, running compliance checks — need to understand document *structure*, not just content. Bounding boxes and typed blocks give agents the primitives to move from "reading" documents to "acting on" them. An agent processing an invoice can distinguish line items from headers from totals because each block is labeled.
Citation-ready output. When every text segment carries a bounding box and confidence score, the path from extracted text back to its source location in the original document is preserved. This is the foundation for citation-ready AI — systems that can not only answer questions but point users to the exact region of the source document that supports each answer.
Mistral has integrated OCR 4 with its Search Toolkit (in public preview), an open-source, composable search framework where OCR 4's structured output supplies citation-ready inputs to the toolkit's ingestion, retrieval, and evaluation workflow.
For teams running a self-hosted AI team, an OCR model that returns structured, confidence-scored blocks is the missing link between raw document storage and a working private knowledge base. When your researcher agent ingests a batch of PDFs overnight — with embeddings computed locally at zero marginal cost — every extracted block lands in a two-layer memory core with its source coordinates intact. No documents leave your VPS, no per-seat SaaS fees accumulate, and every future query can be traced back to the exact region of the original file.
Get OfficeForge — $199Pricing
Mistral's pricing for OCR 4 via the API is $4 per 1,000 pages. The Batch API offers a 50% discount, reducing the cost to $2 per 1,000 pages. Document AI in Mistral Studio is priced at $5 per 1,000 pages and provides a no-code interface to the same engine.
For teams doing the math: at $2 per 1,000 pages via Batch API, processing a 10,000-page document corpus costs $20. For organizations running OCR locally on their own hardware, the per-page API cost drops to zero — only compute and electricity remain.
Accepted Formats
OCR 4 handles common enterprise document formats out of the box: PDF, DOC, PPT, and OpenDocument. No format-conversion preprocessing required — a practical detail that removes a common pipeline bottleneck.
The Bigger Picture
Mistral OCR 4 represents a broader shift in how document AI integrates with the self-hosted AI stack. The era of "OCR as a text-extraction utility" is giving way to "OCR as a structured ingestion layer" — and the teams that benefit most are those building private, citation-aware knowledge systems where documents don't just get *read* by AI, they get *understood*.
The combination of bounding boxes, typed blocks, confidence scores, 170-language support, and single-container deployment makes OCR 4 a credible building block for production document pipelines that need to stay on-premises. Whether you're building a RAG system over internal contracts, an agent that processes multilingual invoices, or a compliance workflow that requires verifiable source citations, the structured output this model provides is what makes those downstream systems actually reliable.
The source material is available at Mistral's original announcement. For teams comparing self-hosted approaches to document AI against SaaS-based alternatives, the economics and architecture choices outlined above are worth a close read.
FAQ
What is Mistral OCR 4?
Mistral OCR 4 is a document-parsing model that extracts text alongside bounding boxes, block-type classifications, and inline confidence scores. It supports 170 languages and can be deployed in a single container on your own infrastructure.
How much does Mistral OCR 4 cost?
The API costs $4 per 1,000 pages. Using the Batch API drops the price to $2 per 1,000 pages (a 50% discount). Document AI in Mistral Studio is priced at $5 per 1,000 pages.
Can Mistral OCR 4 be self-hosted?
Yes. OCR 4 is compact enough to run in a single Docker container, allowing organizations to keep all document data within their own infrastructure for data sovereignty and compliance.
What file formats does OCR 4 support?
It accepts common enterprise formats including PDF, DOC, PPT, and OpenDocument.
How does OCR 4 perform compared to other OCR systems?
Independent human annotators preferred OCR 4 over every leading system tested, with win rates averaging 72%. It also achieved the top overall score (85.20) on OlmOCRBench.