Is Mistral OCR Free or Open Source

Mistral OCR is not free and it is not open source. Mistral OCR 4 is a paid commercial API at $4 per 1,000 pages, $2 per 1,000 through the Batch API, and $5 per 1,000 for Document AI when you want structured fields back. The weights are not openly licensed. The confusion comes from somewhere real: Mistral does publish genuine open weight models under Apache 2.0, and Mistral OCR can be self hosted by enterprise customers. Neither of those makes the OCR product free or open source.

This question comes up constantly, and the answers you find are usually wrong in one of two directions. Some sources say Mistral OCR is open source because Mistral is known for open weights. Others say self hosting means open source. Both are mistakes, and if you are choosing a document pipeline on that basis you will get an unpleasant surprise at procurement. Here is the accurate version.

Is Mistral OCR free?

No. Mistral OCR 4 is a paid API product with per page pricing. The published rates are $4 per 1,000 pages on the standard API, $2 per 1,000 pages if you send work through the Batch API at the 50 percent discount, and $5 per 1,000 pages for Mistral Document AI, which is the annotation layer that returns typed fields instead of plain markdown.

Mistral offers limited free experimentation allowances so you can test the API before committing, which is where a lot of the "Mistral OCR is free" claims come from. That is a trial, not a free tier you can run production document volume on. Once you are processing real invoices at any meaningful rate, you are on the paid rates above.

Rates are from the official Mistral API pricing page, checked August 2026. At half a cent a page the standard rate is genuinely cheap by document API standards, so the honest framing is not "expensive" but simply "paid, and predictable."

Is Mistral OCR open source?

No. Mistral OCR 4 is a proprietary commercial model. Its weights are not published under an open license, you cannot download it from Hugging Face and run it, and you cannot fork or modify it.

The reason this confuses people is that Mistral genuinely is one of the more open weight friendly labs. Pixtral 12B, its general purpose vision model, is released under Apache 2.0 and you really can self host it however you like. Pixtral is not Mistral OCR. It is a general vision language model that can read documents reasonably well, while Mistral OCR is a specialized document model with page layout understanding, paragraph level bounding boxes, typed block classification and inline confidence scores that Pixtral does not provide.

So the accurate sentence is: Mistral publishes open weight models, and Mistral OCR is not one of them.

Is Mistral OCR 4 open source?

Also no, and the version number does not change the answer. Mistral OCR 4 was released on 23 June 2026, replacing OCR 3. It is distributed as a commercial API with enterprise deployment options. Every version of Mistral OCR to date has been a closed weight commercial product.

Can you self host Mistral OCR?

Yes, and this is the part that is genuinely true and often gets misread as open source. Mistral OCR is available to enterprise customers as a single self contained container that runs inside your own infrastructure, so documents never leave your network.

That solves a data residency problem, not a licensing one. You are running a licensed commercial binary on your own hardware under a commercial agreement. You still pay Mistral, you still cannot inspect or modify the model, and you cannot redistribute it. For a hospital, a law firm or a government contractor that cannot send vendor documents to a hosted endpoint, this is often the single feature that decides the purchase, and it is a real advantage over hosted only competitors like Google Gemini, which offers no on premise option at all. It just is not open source.

What are the genuinely free and open source OCR options?

If open source is a hard requirement rather than a preference, there are real options. They are all meaningfully worse at document understanding than a current commercial model, and that tradeoff is the whole decision.

• Tesseract (Apache 2.0) is the long standing default. It reads text well on clean scans and has no concept of document structure. It will not tell you which text was a table row and which was a footer.
• PaddleOCR (Apache 2.0) is stronger on layout and table detection than Tesseract and has good multilingual coverage. Setup is heavier and it expects you to be comfortable with a Python ML stack.
• docTR (Apache 2.0) is a clean, modern detection and recognition library. Good engineering, still a component rather than a document understanding product.
• Pixtral 12B (Apache 2.0) is the closest genuinely open thing to Mistral OCR, since it is a vision language model that will describe and transcribe a page. You give up the bounding boxes, block typing and confidence scores.

The pattern is consistent. Free and open source OCR gives you characters. Paid document models give you characters plus structure. For invoice work the structure is most of the value, because knowing that a number is a line item unit price rather than a page number is what makes the output usable.

Is there a Mistral OCR free API key or free tier limit?

Mistral provides free experimentation access so developers can evaluate the API, with rate and volume limits, and that access is intended for testing rather than production. There is no permanently free production tier for OCR. If you need a real free tier for evaluation at volume, Google's Gemini free tier is more generous, with the significant caveat that Google uses free tier content to improve its products. That matters if your test documents are real vendor invoices.

This is also where metered pricing quietly becomes a budgeting problem rather than a technical one. A flat per page rate like Mistral's is easy to forecast. A token billed model is not, because the bill moves with prompt length and reasoning tokens, and teams that run several of these APIs alongside their cloud infrastructure often find the only way to stay on top of it is to track what the metered spend actually adds up to each month rather than trusting the estimate they built at the start.

How does Mistral OCR pricing compare to the alternatives?

Against the other document APIs, Mistral sits in the middle and its pitch is predictability rather than being cheapest.

Gemini is cheaper per page and currently scores higher on independent community accuracy voting, which is covered in detail in our Gemini OCR vs Mistral OCR comparison. Mistral answers with bounding boxes, confidence scores, a bill you can forecast, and the self hosting option. If you want the full rate breakdown for either one, see Mistral OCR pricing and Gemini OCR pricing. For how Mistral stacks up against Microsoft's document stack specifically, we ran that comparison in Mistral OCR vs Azure Document Intelligence.

What "free" actually costs you on invoice work

The reason this question matters is that people asking whether Mistral OCR is free are usually trying to work out whether they can build invoice extraction without a budget line. You can, and the API bill is rarely what stops them.

Say you process 100,000 invoice pages a year. On Mistral OCR that is $400. On Tesseract it is $0 plus a server. Neither number is what makes the project expensive. What makes it expensive is everything after the text comes back: mapping markdown or raw characters into vendor name, invoice number, date, tax and totals; splitting line items and keeping them attached to the right invoice across page breaks; validating that the extracted line items actually sum to the stated total; handling the vendor who redesigns their template in March; and writing all of it into a file your accounting system will accept.

That is typically several engineer weeks to build and an ongoing maintenance commitment after, and it is identical whether the OCR underneath was free or $5 per 1,000 pages. Choosing the free option saves you the smaller number and leaves the larger one exactly where it was.

The short version

Mistral OCR is a paid, closed weight commercial product at $4 per 1,000 pages that happens to offer enterprise self hosting. Mistral does publish real open weight models, Pixtral 12B among them, but those are general vision models rather than the OCR service. If you need genuinely open source, Tesseract, PaddleOCR and docTR are the honest answers, and you will trade away the document structure that makes commercial models worth paying for.

And if what you actually want is invoice data in a spreadsheet rather than an OCR layer to build on, none of these are the right shape. Upload an invoice at the top of this page and you get the extraction, the field mapping, the line item splitting and the totals validation already done, exported to xlsx, csv, json or QBO. Flat monthly pricing from $49, no per page metering and no schema to maintain. If you were planning to write that pipeline anyway, one of the APIs above is the better buy. If you were not, our invoice data extraction API returns finished invoice fields instead of text you still have to interpret.