Mistral OCR 4 turns document scans into structured data

Mistral AI released Mistral OCR 4 on 23 June 2026. The model does more than transcribe a page: it identifies where text sits, what kind of block it belongs to and how confident it is, giving document-heavy AI systems something closer to a map than a pile of words.

Mistral AI Watch analysis

What happened

OCR 4 supports 170 languages, accepts common office formats and can run in a single container for self-hosted deployments. It is available through Mistral’s API and Document AI, with the company positioning it for enterprise search, retrieval-augmented generation, document extraction and agent workflows.

Key findings

• Layout-aware output
  Bounding boxes and block labels can help downstream systems find tables, equations, signatures and other structures instead of flattening everything into text.
• Confidence included
  Inline confidence scores give verification workflows a useful signal for deciding where a human should check the model’s work.
• Built for multilingual documents
  Mistral says OCR 4 supports 170 languages across 10 language groups, including specialised and lower-resource languages.
• Self-hosting option
  A single-container deployment can keep document data inside an organisation’s infrastructure, which matters for residency and compliance requirements.
• Promising, not magical
  Mistral reports an OlmOCRBench score of 85.20 and a 72% average human-evaluation win rate, but says benchmark results are directional and should be tested on customers’ own documents.

Why it matters

The useful shift here is structural. An OCR system that knows a heading from an invoice total, and can say when it is uncertain, is more valuable to search tools and agents than one that merely produces a clean paragraph.

That could reduce brittle extraction pipelines in legal, financial and administrative work. It also makes the choice between cloud processing and self-hosting more consequential, because the same document workload can be designed around privacy, latency or cost.

Our read

This is a practical upgrade with unusually clear downstream uses, especially where documents are multilingual, visually complex or too sensitive to send elsewhere. Treat the benchmark numbers as a starting signal, then test OCR 4 against the forms, scans and cursed PDFs your organisation actually owns.

What to watch

• Whether independent evaluations reproduce the reported gains on real enterprise documents.
• How well confidence scores identify genuinely risky extraction errors.
• API and self-hosting availability, performance and total cost at production scale.
• Whether agents can reliably act on the added structure without turning one bad bounding box into an administrative adventure.

Discussion spark: Would structured OCR with confidence scores change how your team handles document search, extraction or human review?

not affiliated with or endorsed by Mistral AI