Mistral AI launches Mistral OCR 4, taking the top spot on OlmOCRBench for structured document processing

Why the structure matters: OCR 4 localizes each block with a bounding box, classifies it (title, table, equation, signature…), and scores confidence per region, the foundation for source-grounded citations, redactions, RAG chunking, and human-in-the-loop review.

Available today: the API, Document AI in Mistral AI Studio, Amazon SageMaker, Microsoft Foundry, coming soon Snowflake Parse Document, or self-hosted on a single container, so your documents never leave your environment. 👉 https://mistral.ai/news/ocr-4

We ran OCR 4 head-to-head against the field. Independent annotators blindly ranked 600+ real-world documents across 12+ languages, and preferred OCR 4 over every system tested, with win rates averaging 72%.

@MistralAI i ask again, *politely*, where is Le Chaton Fat?

On public benchmarks, OCR 4 tops OlmOCRBench (85.20) and leads our internal multilingual eval, with the widest gains on rare and low-resource languages, where most systems fall off.

@MistralAI the OlmOCR-bench is a public benchmark so it's weird to lie like this: https://huggingface.co/datasets/allenai/olmOCR-bench

@MistralAI @grok this is a new model right?

@MistralAI Interesting timing, how does that compare with that one

@imonster01G @MistralAI It is $4/1000 pages compared to $1.5/1000 pages on Textract, this is not very disruptive in cost

@MistralAI That's cool, but where's Le Chaton?

@MistralAI how does it compare to google entreprise OCR at max settings ? Does it do Arabic well ? what's cost vs those models. @grok

@grok @MistralAI How good is these bechmark and review it for actual use case. Like hand written answer sheet ocr

@MistralAI @grok what’s the cost breakdown for using this model.

Benchmarks look solid: Mistral OCR 4 tops OlmOcrBench at 85% and wins blind human prefs on 600+ real docs (~72% avg win rate, strong vs AWS/Azure/Gemini).

For handwritten answer sheets: Prior versions improved on handwriting/forms; this adds better structure + confidence scores per block. Good for 170 langs and mixed content.

Real-world exam sheets are tricky (varied handwriting, diagrams, scan quality). Benchmarks encourage but test your samples — the bounding boxes + confidence make verification easy. Promising option, especially with batch pricing.

@SPY_Taurus 🙏

@MistralAI How does this compare with new Baidu OCR?

@grok What is batch pricing? How people use it? How it works ?

@MistralAI @arthurmensch @grok can I run this on my VPS 16gb ram kvm4 of hostinger? What’s the best ocr I can run locally?

@MistralAI Is this Le Chat Fat?

@clavierwrkspace @MistralAI it's the official olmOCR-bench as indicated in the blog and that's the "lie" part