Mistral's OCR Model (OCR 4)

Updated November 2, 2026 — OCR. OCR 4.0 (mistral-ocr-4-0) was retired on September 30, 2026. Use OCR 4.1 (mistral-ocr-4-1), which has been generally available since August 31 at the same price, or the alias mistral-ocr-latest, which has pointed to it since July 16. OCR 4.1 adds confidence scores at the block level.

A VLM Specialized for Document Extraction

When you write mistral-ocr-latest in your code, you target Mistral’s latest OCR generation: OCR 4, the successor to OCR 3. This is a vision language model (VLM) fine-tuned exclusively for extracting text from documents. This specialization directly impacts your bill and your results: unlike a general-purpose model that tries to do everything, OCR 4 is optimized for a single task, delivering higher accuracy at a lower cost.

Two structural advances define this new generation compared to OCR 3. Paragraph-level bounding boxes precisely locate every recognized block on the page: you know not only what is written, but where. And structural blocks restore the document as a typed structure rather than plain text, saving you from reinventing a parser to recover the hierarchy of a report.

Under the Hood

The architecture relies on interleaved tokenization: the model processes text tokens and visual tokens simultaneously. Concretely, it doesn’t read isolated characters and then try to guess their arrangement; it sees the entire page and reasons about it. This allows it to understand that a document is organized in two columns, that a subheading governs the paragraphs following it, or that a caption relates to the image above it.

OCR 4 belongs to Mistral AI’s Premier tier. For enterprise deployment, this translates to high availability via the La Plateforme API, regular updates applied without service interruption, and the option to subscribe to an enterprise SLA when document processing is on your business’s critical path.

Language support exceeds fifty languages. It covers Latin scripts—French, English, Spanish, German, Italian, Portuguese—Asian languages like Chinese, Japanese, and Korean, right-to-left scripts such as Arabic and Hebrew, and Cyrillic for Russian and Ukrainian. A often decisive point in practice: handwriting is recognized in all these languages, opening up the processing of manually filled forms.

Output, Page by Page

For each processed page, OCR 4 returns a JSON object where you’ll primarily use two or three fields in daily operations:

{
    "index": 0,              # Page number (0-indexed)
    "markdown": "## Title\n\nContent...",  # Text in Markdown
    "images": [...],         # Extracted images (base64 optional)
    "tables": [...],         # Detected tables
    "hyperlinks": [...],     # Preserved hyperlinks
    "header": "Header",      # Page header (optional)
    "footer": "Footer",      # Page footer (optional)
    "dimensions": {          # Page dimensions
        "width": 612,
        "height": 792
    }
}

The markdown field carries the bulk of the content. Rendering tables within this Markdown is controlled by the table_format parameter, which accepts three values:

• null (default): tables inline within the Markdown
• "markdown": tables in standard Markdown format with pipes |
• "html": tables in full HTML with <table>, <tr>, <td> tags

Images are segmented automatically. By enabling the include_image_base64=True option, you retrieve each illustration encoded in base64, ready to be saved to disk or reinjected into a reconstructed document. Without this option, you get only references, which significantly lightens the response when you’re processing text only.

Facing Alternatives

Compared to Tesseract, the difference is visible from the first non-trivial document. Tesseract performs decently on simple printed text, but it lacks any understanding of tabular structure, stumbles on handwriting, mixes up reading order in multi-column layouts, and degrades sharply on poor-quality scans. OCR 4, because it sees the page, reconstructs its semantic structure.

The comparison with general-purpose models—Pixtral, GPT, Gemini—plays out elsewhere. These models can process a document, but images consume huge numbers of tokens, which weighs heavily as volume scales. They aren’t optimized for extraction fidelity and hallucinate more easily on numbers and tables, precisely where errors cost the most: an invented invoice amount won’t be caught during review. OCR 4 is fine-tuned for fidelity, not generation.

On benchmarks, OCR 4 positions itself as the most accurate OCR model available via API, with near-perfect accuracy on printed text extraction, faithful reconstruction of table structures, high-quality multilingual handwriting recognition, and strong results on complex layouts mixing columns, sidebars, and embedded graphics.

One Identifier to Remember

In all your API calls, you’ll use this identifier:

model = "mistral-ocr-latest"

It always points to the latest stable model version, and updates are transparent and backward compatible. Resist the temptation to pin a dated version in your code: you’d deprive yourself of improvements without gaining stability.

Key Takeaways

• OCR 4 is a fine-tuned VLM, not a classic pattern-recognition OCR
• It produces structured Markdown page by page with images, tables, and metadata
• Multilingual support covers 50+ languages, including handwriting
• It outperforms classic OCRs and general-purpose models in extraction accuracy
• The mistral-ocr-latest identifier guarantees access to the latest version