The short version: Mistral released the full weights of Mistral Large 4 on October 29, and the license that ships with them is good news for almost every small business. The Mistral Large 4 license lets you use, modify, fine-tune and sell what you build on the model for free. Only one kind of business has to call Mistral first: one that hosts the model and lets outside customers drive it directly. On Mistral’s own API the model costs $0.70 per million tokens in and $2.10 out (read October 31), which works out to about $3.58 a month for a website quote assistant handling 800 inquiries.
That last figure is for a typical three-van trades shop, and we show the math below. The rest of this piece is the license itself, read clause by clause, because the people who will actually feel it are the web agencies and local IT shops that build AI tools for small businesses, and the owners who hire them.
What Mistral released on October 29
Mistral Large 4, nicknamed Le Chonk, first appeared as a hosted preview on October 6. We looked at its office-work claims at the time in our piece on Mistral Large 4 and AutomationBench. On October 29 Mistral published an improved version and, this time, released the weights: the trained model itself, downloadable from Hugging Face, that anyone can run on their own hardware.
It is a big model. The model card describes roughly 1 trillion total parameters, with about 52 billion active for each word it writes, a context window of 1 million tokens, and image input as well as text. Mistral’s announcement says three more weeks of training after the preview pushed it to the front of open-weight models on several coding tests, and that it now offers three reasoning settings (low, medium and high) so you can trade thoroughness against cost.
Those are Mistral’s own benchmark claims, and independent boards will take a few weeks to settle them. The license is different. It is a short document, it is already final, and it decides who can do what with the model starting today.
What does the Mistral Large 4 license allow?
Mistral calls it the Mistral Ecosystem Open-Weight License, and describes it as “a modified version of the MIT license.” MIT is one of the most permissive licenses in software. The grant in this one is just as wide: you may use, copy, modify, publish, distribute, sublicense and sell the model, run and deploy it, train it further, and build derivative works from it. The only standard condition is that the copyright and permission notice travel with any copy you hand on.
Then comes the one modification. If you, or any of your affiliates, operate a “Model as a Service” business, you must sign a separate commercial license with Mistral before using the model, or anything derived from it, for any commercial purpose. Mistral provides a contact form for that request.
Everything turns on that defined term, so it is worth reading slowly. The license defines Model as a Service as giving a third party access to the model for inference or post-training, by API, hosted endpoint or any other means, in a way that lets that third party exercise meaningful control over the inputs, the parameters or the training data. It then carves two things out of the definition: end-user products where the model’s abilities are “solely embedded within specific features or harnesses,” and the simple relaying of requests to a model that someone else hosts.
A separate section adds two more blanket exemptions. Internal use, meaning any use that does not make the model available to third parties, is never caught. Neither is any use through Mistral’s own official products, which includes its API.
Who has to ask Mistral, and who does not
Put those clauses side by side and four common small-business situations sort themselves out.
You use Large 4 inside your own business. Your office manager drafts estimates with it, your bookkeeper summarizes statements, your own staff tool calls it. That is internal use. No commercial license is needed, whether you reach the model through Mistral’s API or a server you control.
You pay Mistral’s API for it. Exempt, because it is use through Mistral’s official products. Most small businesses will meet Large 4 this way, and for them the license changes nothing at all.
Your agency builds a feature on it for your customers. A quote-request assistant on your website, a chatbot that answers service-area questions, an intake form that sorts jobs by urgency. Your customers type into it, but the model sits inside one specific feature with a fixed purpose. That is the “embedded within specific features or harnesses” exclusion, so it is not Model as a Service, and neither you nor the agency needs a separate deal. This is the case the clause seems written to protect: building real products on an open model and selling them.
An IT shop hosts the raw model for its clients. Here the line bites. A managed-service provider that runs Large 4 on its own servers and gives each client an endpoint to send their own prompts, choose their own settings, or fine-tune on their own data is giving third parties meaningful control over inputs, parameters or training data. That is Model as a Service, and it needs Mistral’s commercial license before it earns a dollar from it.
Two details are easy to miss. The affiliate wording means a web agency whose sister company sells hosted AI may be caught across all its commercial use of the model, not only in the sister business. And the license does not define “meaningful control.” A client portal where each customer writes its own system prompt for “their” assistant sits somewhere between a fixed feature and a rented model. If your business model lives in that gap, use Mistral’s contact form and get the answer in writing before you build.
How this compares with the last Mistral Large
The previous flagship, Mistral Large 3, shipped under Apache 2.0, a standard open-source license with no Model as a Service condition at all. OpenRouter’s listing still describes it that way (read October 31). So Large 4 is slightly less open than its predecessor on paper, and the narrowing is aimed squarely at companies that would resell hosting of the model, the kind of business that competes with Mistral’s own API.
For the owner of a small business that is mostly reassuring. The clause protects Mistral’s hosting business, not its claim on what you make. It is also short, which is not true of every license on a giant open model; we read the terms that came with Kimi K3’s weights the same way. Here there is one condition, one defined term and one form to fill in.
What a Mistral Large 4 assistant costs to run
Mistral’s announcement prices the hosted model at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. OpenRouter showed the same rates on Mistral’s endpoints on October 31, against $0.50 and $1.50 for Large 3. That is a 40 percent increase on both sides, for a model Mistral says improved sharply in coding and in how many tokens it needs per task.
Take a plumbing company with three vans whose agency builds a quote-request assistant for its website. Assume 800 inquiries a month, about 4,000 input tokens each once the price list, service area and instructions are included, and about 800 tokens of written reply.
- Input: 800 times 4,000 is 3.2 million tokens, at $0.70 per million, so $2.24.
- Output: 800 times 800 is 640,000 tokens, at $2.10 per million, so $1.34.
- Total: about $3.58 a month. The same workload on Large 3 rates comes to about $2.56.
For comparison, OpenAI’s GPT-6.1 Terra, which we priced at $1 in and $5 out on October 27, comes to about $6.40 for the same 800 inquiries.
One setting matters more than the price card. Mistral’s model card notes that the output limit counts reasoning tokens as well as the visible reply, which means thinking is billed as output, and the default reasoning level is high. If high reasoning adds an assumed 1,500 hidden tokens to each answer, that is another 1.2 million output tokens, about $2.52, and the month rises to roughly $6.10. Mistral’s own guidance is low for quick chat replies and medium for everyday work. A quote assistant is a medium job at most, so ask your agency which level it set.
Caching helps too. The instructions and price list repeat on every request, and repeated input bills at $0.07 instead of $0.70, a tenth of the rate, so a well-built assistant will usually come in under the figures above.
Can a small business run Le Chonk on its own computer?
Not on a desk. The serving command on the model card splits the model across eight GPUs, and the smaller compressed version Mistral offers targets Nvidia’s data-center Blackwell chips. Running the full model yourself is a server-room project, not a back-office one.
The value of open weights for a small business is portability rather than a box in the closet. The same rent-or-own question came up with Meta’s Muse Spark Open. Published weights cannot be withdrawn from whoever builds on them. If Mistral changes its prices or terms, a tool built on Large 4 can move to another host that licenses the model, or to hardware your IT provider runs, without being rebuilt from scratch. Mistral already names Microsoft Foundry and Nvidia-hosted endpoints as places to get it. On October 31 OpenRouter listed only Mistral’s own endpoints, so that choice of hosts is still forming.
The checks worth making this week
If you already use Mistral through its API or its apps, nothing changes for you. If an agency or IT provider is building something for you on Large 4, three questions cover it:
- Where does the model run? Mistral’s API, a licensed cloud host, or the provider’s own servers. Only the last one raises the license question, and only if they offer it to others as a service.
- Is it a feature or a rented model? A fixed tool with a defined job is the embedded-feature case. If you get your own prompt box and settings on a model they host, ask whether they hold Mistral’s commercial license.
- What reasoning level is set? High is the default and adds about 70 percent to the bill in our example. Medium or low is right for most customer-facing work.
The license also keeps the usual “as is” wording: no warranty, and Mistral is not liable for what the model or its output does. That is standard for open software and it carries a plain message. The model is a capable tool you now get to own outright, and you, not Mistral, stand behind every answer it gives your customers. Keep a person on the approvals that matter.
Frequently Asked Questions
Is Mistral Large 4 free for commercial use?
Yes, for most businesses. The Mistral Ecosystem Open-Weight License lets you use, modify, fine-tune and sell products built on the model free of charge. The one exception is a business that operates a Model as a Service offering, meaning it hosts the model and gives outside customers meaningful control over its inputs, settings or training. That business needs a separate commercial license from Mistral first.
Does a chatbot on my website count as Model as a Service?
Not if the model sits inside a specific feature with a fixed job, such as a quote assistant or a service-area chatbot. The license excludes end-user products where the model’s abilities are embedded within specific features or harnesses. It also exempts any use through Mistral’s own API and any internal use that does not make the model available to third parties.
How much does Mistral Large 4 cost per month for a small business?
Mistral prices the hosted model at $0.70 per million input tokens and $2.10 per million output tokens (read October 31, 2026). A website quote assistant handling 800 inquiries a month at about 4,000 input and 800 output tokens each costs about $3.58. Leaving the reasoning level on its default of high can add about 70 percent to that, so set it to medium or low for routine work.
Can I run Mistral Large 4 on my own computer?
Not on ordinary office hardware. The model has about 1 trillion parameters, and the serving setup on its model card spreads it across eight data-center GPUs. Small businesses will normally use it through Mistral’s API or a cloud host, and the open weights matter mainly because a tool built on them can move to another licensed host later.
If you have had an agency or IT provider build an AI tool for your business, do you know whose servers it runs on, and who could change its terms tomorrow?
