Is Mistral OCR Open Source? Weights, License and Self-Hosting Options

The short answer: no. Mistral OCR is not open source. Mistral publishes open weights for several of its text models, but the OCR model is not one of them. There is no downloadable checkpoint, no Hugging Face repository from Mistral for it, and no license that lets you run it on your own hardware. Mistral OCR is available only as a paid API.

This confuses people for a good reason. Mistral built its reputation on open weights. Mistral 7B and the Mixtral models were released under Apache 2.0 and became the default starting point for anyone who wanted a capable model they could actually host. So when Mistral shipped a document model that beat much of the market on layout and tables, a lot of teams assumed the same thing would happen and started looking for the repository. It is not there.

Is Mistral OCR open source?

No. The OCR model is API-only. You call it through Mistral's platform using the model alias mistral-ocr-latest, or the pinned snapshot id mistral-ocr-4-0 for the current version, Mistral OCR 4. What Mistral distributes is access to the endpoint, not the weights behind it.

Searching Hugging Face for it will turn up results that look promising and are not. You will find community Spaces that wrap Mistral's OCR API behind a web form, which still bill against an API key. You will find fine-tuned Mistral text models that people trained to clean up OCR output, which are a different thing entirely. Neither is the OCR model. Mistral's own model-weights documentation lists what has been released, and the OCR model is absent from it.

Why does it matter whether the weights are published?

Three reasons, and only one of them is cost.

The first is data residency. If your documents cannot leave your own infrastructure, an API is not an option no matter what it costs. This comes up constantly in healthcare, defense work and anything touching a signed data processing agreement that predates generative AI. An open model you host is the only shape that fits.

The second is the retention default. Mistral's platform limits page states that uploaded files are retained for 30 days unless deleted earlier. That is a reasonable default and you can delete files through the API as soon as processing finishes, but it is a thing you now have to actively manage. A model running on your own hardware has no such question attached to it.

The third is unit cost at volume, and this is where most people's intuition is wrong. Self-hosting is not automatically cheaper. It is cheaper only above a break-even point that is much higher than teams expect, because a rented GPU bills for every hour it exists, not for the pages you push through it.

Can you self-host Mistral OCR?

Not the model itself. If self-hosting is a hard requirement, the practical answer is to use a different model rather than to wait for Mistral to publish this one.

The closest genuine substitute is olmOCR from the Allen Institute, released under Apache 2.0. It needs about 12 GB of VRAM and 30 GB of disk, and its authors report processing a million PDF pages for roughly $176 in compute. Be careful how you read that number, though: it is the compute cost at high utilization, and it excludes the engineering time to run the thing. olmOCR also returns markdown rather than named fields, so if you need an invoice number in a database column, the extraction step is still yours to build on top.

We worked the arithmetic out properly in self-hosted OCR cost, and the shape of it is worth knowing before you commit. An AWS g6.xlarge with an L4 runs about $587 a month on demand. Push 10,000 pages a month through it and you are paying roughly $58 per 1,000 pages, which is around forty times what a cloud API charges. Push 250,000 and you are at about $2.35. Break-even against a $1.50 API rate lands near 390,000 pages a month. Below that, the API wins on pure cost, and it is not close.

Is Mistral OCR free?

No, though you can test it without paying. Mistral gives new accounts trial credits, which is enough to run real documents through the model and judge the output. After that it is billed per page. Mistral OCR 4 is $4.00 per 1,000 pages and Mistral Document AI, which returns JSON fields against a schema you define, is $5.00 per 1,000 pages. Batch processing halves both.

There is one wrinkle worth knowing, because it is on Mistral's pricing page and almost nobody reports it: Mistral publishes a second, cheaper OCR rate of $3.00 per 1,000 pages that belongs to a different product. It applies to documents you ingest into Libraries, Mistral's document store built so that agents can query your data, and it comes with two extra meters attached, one for indexing and one per retrieval call. It is not a cheaper way to run a plain extraction batch. We break the whole rate card down on our Mistral OCR reference, including which line you actually get billed on.

That agent-shaped use case is where a lot of document work is heading, and if you are assembling that kind of stack out of ready-made parts instead of writing it yourself, there is now a whole market of prebuilt AI agents to start from rather than a blank file.

What do you get from the API that you would have to rebuild yourself?

More than the phrase "just OCR" suggests, and this is the honest argument for staying on the API. Mistral OCR 4 returns markdown per page with reading order preserved, tables as markdown or HTML, detected hyperlinks, header and footer content on request, and paragraph-level bounding boxes with one of thirteen block labels including table, equation and signature.

It also returns confidence scores, which is the feature most worth protecting. Setting confidence_scores_granularity to "word" gives you a value for every word on the page and in every table cell. That number is what lets you route the uncertain few percent of a batch to a person instead of either checking everything by hand or shipping silent errors. A general chat model used as an OCR engine gives you no equivalent, which is the real difference between LLM OCR and a purpose-built document model.

Is Mistral OCR 4 open source?

No, and version 4 is if anything less likely to be released than earlier ones. It is Mistral's current commercial document model, badged as new on the pricing page, and block extraction is gated to it specifically. Older snapshots accept the include_blocks parameter and return an empty array rather than an error, which tells you Mistral is using version gating to differentiate the paid product. That is not the behavior of a model heading for an open release.

So what should you actually do?

Decide which constraint is real, because they point in different directions.

If documents genuinely cannot leave your network, stop evaluating APIs and go straight to an open model on your own hardware, with olmOCR as the sensible default and the understanding that field extraction remains your problem. If your volume is above roughly 400,000 pages a month and steady, run the numbers on both, because at that point self-hosting starts to win on cost alone. If neither applies, which covers most teams, the per-page API is cheaper, better and available this afternoon, and the interesting question is not whether the weights are open but whether the fields land correctly in your system.

That last part is where document projects actually fail. A model that reads a page well is necessary and not sufficient. Between reading the page and having trustworthy data in your accounting system sit classification, validation, a review queue for the low-confidence values, export and retries, and on most projects that surrounding work costs considerably more than the recognition API ever will. Whether you buy the model or host it, that is the part to plan for.