Mistral OCR 4

473 points · 128 comments on HN · read original →

Points and comments are a snapshot, not live.

Mistral OCR 4 achieves top human-preference and benchmark scores for document extraction.

Mistral OCR 4 adds bounding boxes, block classification, and confidence scores. It supports 170 languages, runs in a single container for self-hosted deployment, and costs $4 per 1,000 pages ($2 batch). On OlmOCRBench it scores 85.20 (first), on OmniDocBench 93.07. Independent annotators preferred OCR 4 over all competitors in head-to-head tests, with an average 72% win rate. The model integrates with Mistral's Search Toolkit for RAG pipelines.

Document AI (same engine, additional parameters) returns structured JSON per user schema and costs $5 per 1,000 pages. Available via API, Mistral Studio, Amazon SageMaker, and Microsoft Foundry. Self-hosting is available for enterprise customers.

What commenters are saying

Commenters generally praise Mistral OCR 4's performance on degraded documents and handwritten forms, with several noting it surpasses classic tools like Abbyy FineReader. One user successfully processed 55-year-old paper files. Another reports high accuracy digitizing handwritten forms with minimal human review needed.

Some skepticism arose about using LLMs for OCR on dates from receipts; one commenter reported a 20% error rate with Opus 4.8 despite high-confidence ratings. A LlamaIndex contributor announced plans to benchmark against their ParseBench benchmark. Two users noted competing OCR announcements from Baidu, and a discussion about SSL certificate validity was resolved with a root certificate chain analysis.