Mistral OCR 4 Brings Bounding Boxes, 2,000 Pages Per Minute to Enterprise Document AI

Quick Facts

• Mistral OCR 4 processes up to 2,000 pages per minute on a single GPU and supports 170 languages across 10 language groups.
• The model scored 85.20 on OlmOCRBench and achieved a 72% average win rate over competitors in a blind evaluation of 600-plus real-world documents.
• OCR 4 is available through Microsoft Foundry and can be deployed entirely within a customer’s own infrastructure.

Mistral AI released OCR 4 on June 23, its fourth-generation document intelligence model and the most technically ambitious yet. The Paris-based company is positioning the release as a direct play for regulated enterprises that need structured data extraction without sending sensitive documents to third-party cloud services.

The central change from previous versions is structural output. Earlier OCR systems, including prior Mistral generations, return a flat stream of extracted text. OCR 4 returns a layered representation where every block carries a bounding box, a content type classification, and a confidence score at both the page and word level.

Bounding boxes were Mistral’s most-requested feature from enterprise customers. Without location data, downstream systems cannot trace an extracted value back to a specific spot on a specific page, forcing engineering teams to build and maintain that mapping layer themselves. OCR 4 includes it by default.

Confidence Scores Drive Automation at Scale

The confidence scoring system lets organizations build automated review workflows without human review on every page. High-confidence extractions can be approved automatically. Low-confidence regions route to human reviewers. That architecture is critical for high-volume use cases in legal, financial, and healthcare settings where manual review of every document is not feasible.

Block classification covers body text, titles, lists, tables, images, equations, captions, code, references, side notes, headers, footers, and signatures. Each entry includes coordinates and extracted content.

The model accepts PDF, DOC, PPT, and OpenDocument formats. On Mistral’s internal multilingual benchmark, OCR 4 leads across all eight language groups tested, with the largest performance gap appearing in specialized and low-resource languages where competing systems degrade the most.

Benchmark Results and Competitor Comparisons

Mistral reports OCR 4 scored 85.20 on OlmOCRBench, the top result on that public leaderboard, and 93.07 on OmniDocBench. The 72% average win rate came from an independent annotator evaluation across more than 600 documents in over 12 languages sourced from third-party vendors.

Mistral also disclosed specific scoring artifacts it found during benchmarking, including ground-truth errors in reference annotations and LaTeX notation mismatches. The company said it treats aggregate scores as directional rather than definitive, an unusual transparency move for a product launch.

Early production users reported sharp performance gains against incumbent tools. Rogo, a financial AI platform, benchmarked OCR 4 against leading agentic document parsers on a chart and figure-dense financial dataset. “We reached equivalent accuracy at roughly 8x lower cost and 17x lower latency,” said Aidan Donohue, AI engineer at Rogo. “For production use cases at scale, that delta compounds fast.”

Anaqua, an intellectual property management company, tested OCR 4 on high-volume patent docketing workflows. “Mistral OCR is roughly 4x faster per page than our incumbent provider,” said Ivan Mihailov, AI engineer at Anaqua.

Microsoft Foundry Integration and Enterprise Go-to-Market

OCR 4 is available now through Microsoft Foundry alongside Mistral Medium 3.5, giving Microsoft’s enterprise customer base direct access through a platform they already use. Mistral CRO Marjorie Janiewicz said the integration is designed to help organizations “move beyond text extraction and unlock structured, actionable data that powers automation, compliance, and AI-driven decision making at scale.”

The Microsoft channel matters because it lowers the procurement and integration barrier for large enterprises that have already standardized on Azure. Mistral does not have to win those IT procurement battles independently.

OCR 4 is the fourth Mistral document model in roughly 15 months. The original launched in March 2025. OCR 3 followed in December 2025 with a reported 74% win rate over its predecessor. Mistral has scheduled a live webinar with demos on July 7 at 6:00 p.m. CET for prospective customers evaluating the model in production contexts.

Read more: Mistral launches OCR 4, turning document extraction into a full enterprise AI play