Mistral releases OCR 4 with bounding boxes, block classification, and 170-language support

Mistral AI released OCR 4 (mistral-ocr-4-0) on June 23, 2026, its latest document understanding model. The update adds structured spatial output — bounding boxes, content-type labels, and per-word confidence scores — that goes significantly beyond plain text extraction. The mistral-ocr-latest alias now resolves to OCR 4.

What's new

The headline addition is a new include_blocks parameter. When enabled, each page response includes a blocks array with paragraph-level bounding boxes and a structural label for every content element. According to Mistral, "OCR 4 returns bounding boxes, typed-block classification (titles, tables, equations, signatures, and more), and inline confidence scores."

Capabilities at a glance:

• Bounding boxes: Every text block is localized with coordinates
• Block classification: Content labeled as title, table, equation, signature, or body text
• Confidence scores: Per-word and per-page confidence returned inline, not just a single document-level value
• Language breadth: 170 languages across 10 language groups, including rare and low-resource languages
• Self-hosted deployment: Single-container packaging for on-premises data residency requirements

Pricing:

• API: $4 per 1,000 pages
• Batch API: $2 per 1,000 pages (50% discount)
• Document AI: $5 per 1,000 pages

Accepted formats include PDF, DOC, PPT, and OpenDocument files. The pages parameter now also accepts comma-separated ranges such as "0,2-4", in addition to the previous list format.

Context

Mistral has been building its OCR product into a standalone document intelligence offering over the past year, distinct from its language model lineup. The focus on spatial metadata — coordinates, type labels, confidence — positions OCR 4 against structured document extraction services from Google Document AI and Amazon Textract, not just generic text extraction.

The single-container self-hosted option addresses a gap that has kept regulated industries from using cloud OCR APIs: document data need not leave the customer's own infrastructure. Mistral Search Toolkit integration is also supported for RAG and enterprise search pipelines.

Why it matters

Flat text extraction is rarely what production document pipelines actually need. Knowing where a table sits on a page, what type a block is, and how confident the model is at the word level enables downstream systems to preserve document structure, filter noise, and flag ambiguous regions for review — capabilities that matter in financial document processing, legal discovery, and scientific literature ingestion.

The 170-language support, with explicit coverage of rare and low-resource languages, expands the model's usability well beyond English-first document workflows. For teams already using Mistral's language models, the shared API surface and RAG toolkit integration lower the friction of adding structured document parsing to an existing pipeline.