Mistral OCR

Current version is Mistral OCR 3 (mistral-ocr-2512), released December 18, 2025 — the first version of this product that genuinely competes on pricing with the hyperscaler document-AI services. Where AWS Textract charges ~$65/1,000 pages for forms-and-tables extraction and Google Document AI charges $30–45/1,000, Mistral OCR 3 is $2/1,000 pages standard or $1/1,000 via the Batch API. The trade-off is that this is a proprietary closed API without SOC 2 / HIPAA / FedRAMP compliance attestations published, and without custom model fine-tuning — so for regulated industries that need a compliance paper trail, AWS/Google/Azure still win on procurement posture even when Mistral wins on price.

What Mistral OCR 3 does well: handwriting recognition (88.9% accuracy per independent benchmarks, versus Azure's 78.2% and DeepSeek's 57.2%), complex table reconstruction with HTML output preserving row/column spans, mathematical expressions in LaTeX, and multilingual document parsing across 100+ languages. Independent reviews note it's weaker on extremely complex multi-column layouts (magazine-style) where it sometimes imposes table structure on columnar text. For financial data extraction specifically, human-in-the-loop review is still recommended — structural fidelity can mask individual-digit OCR errors that look correct but aren't.

Self-hosting: Mistral offers enterprise self-hosting for organizations with classified or highly sensitive data needs, but this is a separate commercial track — not an open-weightA model where the trained weights are freely downloadable — you can run it yourself without contacting the creator. Llama, Mistral, Qwen, and Gemma are open-weight. Open-weight does not mean open-source: the training data and code often stay private. The license still governs what you can do with the weights, including whether you can use them commercially. release. Contact Mistral sales if that's your deployment shape.