0
MODEL SIGNAL · MISTRAL

OCR 4

A specialized document extraction model providing structured outputs with layout analysis and confidence scoring across 170 languages.

CATEGORYMultimodal
Key Features
  • Structured document extraction
  • Bounding box generation
  • Block classification
  • Inline confidence scores
  • Supports 170 languages

Provider announcement →

Read the Model Signal report →

OCR 4

Identity

OCR 4 is a multimodal document intelligence model released by Mistral in June 2026 (sn_wire_item:d5f0c082-453b-4b10-9f49-47562c190bc0). It represents the fourth generation of the provider's OCR technology and is designed to provide structured document extraction across 170 languages (sn_model_face:1777050d-3850-472a-95b9-1e428570676b; sn_wire_item:9a55a272-9614-4d2f-989f-fc25761a712a).

What it is

OCR 4 is a specialized model that converts unstructured documents into structured representations by performing layout analysis and block classification (sn_model_face:1777050d-3850-472a-95b9-1e428570676b). Unlike traditional OCR utilities that focus solely on raw text extraction, OCR 4 provides a document intelligence layer that includes spatial bounding boxes and per-word confidence scoring (sn_wire_item:9a55a272-9614-4d2f-989f-fc25761a712a). It is intended to transform PDFs and other document formats into data suitable for direct database ingestion, facilitating the automation of complex document intake (sn_wire_item:d5f0c082-453b-4b10-9f49-47562c190bc0).

Capabilities & benchmarks

The following facts are reported by the model provider (sn_model_face:1777050d-3850-472a-95b9-1e428570676b):

  • Provider: Mistral
  • Category: Multimodal
  • Key Features: Structured document extraction; Bounding box generation; Block classification; Inline confidence scores; Supports 170 languages.

Additional capabilities reported by news sources include:

  • Release Date: June 23, 2026 (sn_wire_item:d5f0c082-453b-4b10-9f49-47562c190bc0).
  • Structured Output: The model returns structured representations of entire documents, moving beyond raw text to include block-type classification (sn_wire_item:9a55a272-9614-4d2f-989f-fc25761a712a).
  • Data Quality: The model is designed to turn unstructured PDFs into "near-database inputs" by providing per-word confidence scores (sn_wire_item:d5f0c082-453b-4b10-9f49-47562c190bc0).

How it compares

OCR 4 is described as moving OCR from a simple utility to a full enterprise document-intelligence layer (sn_wire_item:9a55a272-9614-4d2f-989f-fc25761a712a). It is positioned as a replacement for manual document intake or brittle, hand-built OCR pipelines used for processing invoices and contracts (sn_wire_item:d5f0c082-453b-4b10-9f49-47562c190bc0).

Where it fits

The model is targeted at enterprise AI applications requiring high-quality document extraction, particularly in the finance, logistics, and public