Mistral Launches OCR 4 for Document Intelligence and Enterprise RAG

Mistral AI released OCR 4 on June 23, 2026, a small, focused model that runs in a single container for fully self-hosted deployments. [3] The service powers Mistral's Document AI stack with native paragraph-level bounding box extraction. [5] Alongside extracted text, the model returns bounding boxes, inline confidence scores, and block classification for elements including titles, tables, equations, and signatures. [3] Self-managed deployment is available to enterprise customers, keeping document data within their environments for residency, sovereignty, and compliance. [3] Bounding boxes localize text for in-context highlighting and reliable data pipelines. [3]

On public benchmarks, OCR 4 achieved the top overall score of 85.20 on OlmOCRBench. [3] Mistral states that independent annotators prefer OCR 4 over every leading OCR and document-AI system tested, with win rates averaging 72 percent. [3] The model supports 170 languages across 10 language groups, delivering measurable gains on specialized and low-resource languages where several competing systems degrade. [3] These block types and inline confidence scores drive source-grounded citations, redactions, and human-in-the-loop verification. [3] The structured output supplies citation-ready inputs to ingestion, retrieval, and evaluation workflows for RAG and enterprise search. [3]

OCR 4 operates as an ingestion component of Search Toolkit, which is Mistral's open-source, composable search framework that was announced at the AI Now Summit. [3] For developers and enterprises, the OCR 4 model is priced at $4 per 1,000 pages, while the cost increases to $5 per 1,000 annotated pages. [5] Users can run OCR at scale via Mistral's Batch API and access official cookbooks for tasks like healthcare document processing, document chunking, and extracting data via annotations. [4] Additionally, the model supports cost-efficient, high-throughput batch processing. [3]

Mistral’s launch of OCR 4 emphasizes a clear transition from basic text extraction to citation-ready data pipelines tailored for enterprise search and RAG workflows. By offering self-managed, single-container deployments alongside native bounding box capabilities, the company addresses strict corporate requirements for data residency and compliance. The model’s competitive benchmark scores suggest it can challenge existing document intelligence solutions on complex document layouts. The pricing model separates standard OCR from structural annotations, giving developers cost flexibility based on specific operational needs. What the sources don't address: How OCR 4’s latency scales under heavy concurrent request loads in large enterprise environments.