The new OCR model returns markdown-structured text with bounding boxes, typed blocks, and per-word confidence across 170 languages — and can run on a single container on-prem.
2026-06-23 · by Moe Ameen
Mistral released OCR 4 on June 23, 2026, the latest version of its optical character recognition and document-understanding model. It ships as the model id `mistral-ocr-4-0`, and Mistral's `mistral-ocr-latest` alias now points to it. The pitch is document intelligence: turn PDFs, slide decks, scans, and images into clean, structured, machine-readable text instead of a flat character dump.
The new version returns bounding boxes that localize each element, typed-block classification that labels titles, tables, equations, and signatures, and confidence scores reported per page and per word, with the extracted text formatted as markdown. It supports 170 languages across 10 language groups, with Mistral reporting gains on rare and low-resource languages. On public benchmarks the company reports 85.20 on OlmOCRBench and 93.07 on OmniDocBench, and says independent annotators preferred OCR 4 over competing systems at an average 72% win rate across more than 600 multilingual documents.
OCR 4 is available through the Mistral API at $4 per 1,000 pages, or $2 per 1,000 pages via the Batch API, alongside the no-code Mistral Studio, Amazon SageMaker, and Microsoft Foundry, with a Snowflake Parse Document integration listed as coming soon. Mistral also offers a self-hosted deployment for enterprise customers, saying the model is compact enough to run on a single container for data residency, and positions OCR 4 as an ingestion component for RAG and enterprise search, including in its open-source Search Toolkit.
Here is how to act on this today. The reason OCR 4 matters to a creator is not the benchmark scores — it is that documents you have been ignoring are now trivial to mine. Run a freshly published industry report, a competitor's whitepaper, or your own slide deck through OCR 4, get back clean markdown, and drop it into Kompozy as a source. In one pass Kompozy turns that single document into a carousel breaking down the findings, a LinkedIn post and an X thread in your voice via the Persona Brief, a blog article, and an email newsletter, then schedules and publishes them across your nine connected platforms.
The timely play is speed. When a notable report or document drops, the creators who comment on it first capture the reach. OCR 4 collapses the read-and-summarize step to seconds; Kompozy collapses the write-and-publish step to a single approval. You can react to a document the day it is published instead of three days later when the conversation has moved on. The launch itself is content too — your take on what better document OCR means for your niche can be the first thing you ship with this workflow.
Mistral released OCR 4 on June 23, 2026 as the model id mistral-ocr-4-0. Its mistral-ocr-latest alias now points to the new version.
Structured output — bounding boxes, typed-block classification (titles, tables, equations, signatures), and per-word confidence scores — with markdown formatting, 170-language support with low-resource gains, and a single-container self-hosting option, plus reported benchmark gains over prior systems.
The API is $4 per 1,000 pages standard and $2 per 1,000 pages via the Batch API, and Mistral's Document AI product is $5 per 1,000 pages. A self-hosted option is available to enterprise customers. Check Mistral's pricing page for current rates.
Use it to extract clean, structured text from reports, decks, and scans, then feed that text into a content engine like Kompozy to generate and publish carousels, blogs, newsletters, and platform-native posts across nine platforms.