Mistral OCR

Mistral OCR is Mistral AI’s document-understanding API — designed to read PDFs, screenshots, and image-heavy documents with state-of-the-art layout preservation, equation parsing, table reconstruction, and multilingual coverage.

Hosting model — Mistral partnership

The current generation (Mistral OCR 3 / mistral-ocr-2512) is proprietary and not open-weight — distributed via Mistral’s la Plateforme API. We surface it as a hosted endpoint on ScaLabs Cloud through a partnership with Mistral: same OpenAI-compatible interface as our other utility APIs, backend routed through Mistral’s infrastructure under a B2B agreement.

The earlier Mistral OCR-25-03 has weights available on request from Mistral. If your workload requires fully on-shore Kathmandu hosting (no data transit to Mistral’s infrastructure), we can host OCR-25-03 directly — ask us.

When to pick Mistral OCR over GLM-OCR

• Complex layouts. Multi-column academic papers, financial reports, forms with mixed text-and-table content. Mistral OCR reconstructs the reading order more reliably.
• Equations and tables. SOTA on LaTeX equation extraction and table structure preservation.
• You’ll tolerate a small upcharge. Mistral OCR is roughly 2× the price of GLM-OCR per page; the accuracy uplift on complex documents justifies it for many enterprise workloads.

When to pick GLM-OCR instead

• High volume, simple documents. Receipts, invoices, screenshots — GLM-OCR is half the price and accurate enough.
• Data-residency requirements that forbid transit to Mistral. GLM-OCR runs entirely on our Kathmandu hardware; Mistral OCR 3 transits through Mistral’s infrastructure (unless you arrange the 25-03 on-shore setup).
• Simple text extraction with no layout requirements.

Pricing

EUR 0.001 per page. Flat. Billed against your tenant card on file; no deposit required.

Limits

• Per-tenant rate limit: 30 pages per second
• Image size limit: 50 MB per page (PDFs split automatically)
• Supported formats: png, jpg, webp, pdf, tiff

Output formats

Same as GLM-OCR: raw text, markdown with preserved structure, JSON with a caller-supplied schema, or interleaved text + bounding boxes for audit.