Mistral released OCR 4 on 23 June 2026, the latest version of its optical character recognition model, via its official announcement.

The model extends the core capability Mistral established with its original OCR release: multimodal understanding of text, tables, equations, and embedded images within documents. That combination distinguishes document-understanding models from basic text-extraction tools. Where a standard OCR engine simply converts pixels into characters, Mistral's approach treats a document as a structured object with semantic meaning — parsing layout and content relationships simultaneously rather than processing them separately.

This foundation already operates at scale. When Mistral launched its first OCR model in March 2025, it became the default document-understanding tool in Le Chat, Mistral's consumer and professional assistant, reaching millions of users in a single deployment. OCR 4 builds on that existing user base.

Document processing has become a central focus in enterprise AI. The quality of retrieval-augmented generation (RAG) systems — which feed documents into AI models for reasoning and synthesis — depends directly on input quality. If tables are incorrectly parsed, equations removed, or multi-column layouts flattened into single lines of text, the AI's reasoning suffers regardless of how capable the model itself is. High-quality document understanding is therefore essential to reliable production AI, not an afterthought. Mistral's commitment to iterating on this component — now reaching a fourth named version — reflects how vital structured document ingestion has become for real-world AI deployments.

In this author's view, the Le Chat default-model strategy deserves as much attention as the model itself. Routing millions of actual document-processing tasks through OCR gives Mistral a feedback channel that API-only competitors do not have. The diversity of real usage — across languages, document formats, and specialized vocabularies — surfaces edge cases that benchmark tests typically miss. This distribution advantage compounds over time.

For teams building document-processing systems into their infrastructure, the jump to version 4 signals that Mistral intends to keep improving this capability rather than freezing it. Full technical specifications — performance benchmarks, how the model handles long documents, and any improvements to equation or table processing — are available in the announcement.