OCR 4, Moderation 2, and Leanstral
On this page
Niche Specialists
Beyond code and audio, Mistral maintains models built for narrow tasks: extracting document content, securing generated content, and writing formal proofs. Each complements the generalists rather than competing with them. This lesson also clarifies the status of the Magistral family, whose name you may still encounter in older tutorials.
OCR 4 — Document Extraction
OCR 4, released on June 23, 2026, as version 4.0 on the Premier tier, is Mistral’s optical character recognition service and replaces OCR 3. Calling it a “text reader in images” would be reductive: what it produces is not a string of words, but a representation of the document.
It first understands structure: titles, paragraphs, lists, and tables are typed as distinct elements, directly usable for downstream processing. It then provides bounding boxes at the paragraph level—that is, the exact coordinates of each block on the page—which lets you highlight a clause on screen, annotate a document, or have an operator visually verify an extraction instead of asking them to take the result on faith. It extracts tables while preserving row and column organization, handles multi-page documents without pagination limits, and processes forms by restoring fields with their values.
On the input side, OCR 4 accepts a wide range of formats: PDF, PNG, JPEG, and TIFF images, scanned documents, and smartphone photos. Extraction quality holds up on degraded sources, skewed scans, or low resolution; in digitization projects, the documents to process are rarely the best-scanned ones in the batch.
Common use cases follow this logic. Digitization handles invoices, contracts, and purchase orders. Regulatory compliance relies on automatic data extraction from legal documents. Intelligent archiving makes a historical document collection indexable. Finally, a data pipeline can populate a database from paper documents without manual re-entry.
OCR 4 is rarely integrated alone. The most common combination chains three models:
- OCR 4 extracts structured content from a PDF contract
- Mistral Small 4 analyzes this content and produces a structured summary
- Mistral Embed generates vectors for indexing in a vector database
Magistral — Deprecated Family
Models Removed
The Magistral family (Medium and Small), dedicated to reasoning, is deprecated: all its variants were gradually removed between late 2025 and mid-2026. Reasoning is now integrated into the generalists: Mistral Small 4 unifies instruct, reasoning, and code; Mistral Medium 3.5 takes over for the most demanding reasoning. For details on this convergence, refer to the New Features course.
If you resume an existing project or follow a tutorial that still calls Magistral, migration depends on the task’s demands. For routine reasoning—logic, standard mathematics, analysis—switch to Mistral Small 4 and its integrated reasoning mode. For demanding reasoning, complex demonstrations, or multi-step planning, move to Mistral Medium 3.5. In both cases, plan to rerun your test sets: outputs change form even when conclusions remain identical.
Mistral Moderation 2 — Content Safety
Mistral Moderation 2, released in March 2026 as version 26.03 on the Premier tier, focuses on detecting problematic content and securing AI applications. It classifies content by identifying hate, violence, sexual material, or danger, and it detects jailbreak attempts—that is, messages constructed to make a model bypass its own safety guidelines. It works in many languages, which avoids having to translate before moderating. It also accepts custom guardrails: you define your own rules for your agents and conversations, beyond the categories provided by default.
Its typical use involves installing it around a generalist model rather than inside it. Upstream, it filters user requests before they reach the main model. Downstream, it checks responses before they go back to the user. Used for monitoring, finally, it analyzes past conversations to detect abuse. On a service open to the public, all three positions combine: the first protects your model, the second protects your users, the third protects you.
Leanstral 1.5 — Formal Proofs
Leanstral 1.5, released on July 2, 2026, is the most specialized model in the entire ecosystem. It is an open source foundation model designed for Lean 4, the language of formal mathematical proofs.
Its audience is narrow and perfectly identified: mathematics researchers, formal verification teams for critical software, and proof library developers. Its open source release matters to them because it allows labs to host it and adapt it to their own formalisms.
Key Takeaways
- OCR 4 (June 2026) is a complete document understanding service: text, structure, bounding boxes by paragraph, and typed structural blocks
- Magistral is deprecated: reasoning now goes through Small 4 (routine) and Medium 3.5 (demanding)
- Moderation 2 ensures AI application safety with jailbreak detection and custom guardrails
- Leanstral 1.5 (July 2026) is an open source foundation model for formal proofs in Lean 4
- These specialists combine with each other and with generalists in multi-model pipelines