Breaking

Mistral Releases OCR 4 Document Model With Structured Output

June 23, 2026 at 12:34 EDT

Foundation Models
Sales & Business
AI Agents

Mistral AI formally announced its document-recognition model "Mistral OCR 4" on June 23, 2026, making it available the same day. In addition to text extraction, the new model returns structured output including the position of each block (bounding boxes), typed-block classification, and confidence scores, and it supports 170 languages. According to the official announcement, OCR 4 is positioned as a document ingestion component for RAG and search pipelines.

June 23, 2026 · Mistral AI

Mistral OCR 4 tops the document-parsing benchmark — and adds the structure enterprises asked for

The new release posts a leading 85.20 on OlmOCRBench while adding bounding boxes, typed-block classification and confidence scores — turning raw text extraction into production-ready structured output.

85.20

OlmOCRBench — top score

93.07

OmniDocBench score

72%

Avg. human-eval win rate vs rivals

170

Languages (10 groups)

OlmOCRBench: the climb across the series

From the original Mistral OCR to OCR 4 — score built to scale.

~72

Original
Mistral OCR

→

85.20

OCR 4
(top score)

Versus agentic parsers (financial QA dataset)

Each block = 1× of the agentic-parser baseline.

4×

faster per page

8×

lower cost (equal accuracy)

17×

lower latency

The headline feature — structured output

Bounding boxes Typed-block classification Confidence scores Titles · tables · equations · signatures

Enables citation generation, redaction, human-in-the-loop verification and semantic chunking. Self-hostable in a single container — documents never leave the organization (a draw for GDPR-focused European firms).

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.