Mistral Large 4 for Business: What Europe's New Model Changes

On October 6, 2026, Mistral AI unveiled Mistral Large 4, a model of roughly one trillion parameters, currently available as a preview through its API, with the weights promised for public download by the end of the month. For businesses, it is the first time a European model of this caliber can be hosted on infrastructure of their choosing, without relying on an American or Chinese provider. Between the announcement and a production deployment, though, there are questions about cost, licensing and hardware worth answering before you switch.

This article covers what the model actually delivers, what it costs, who can realistically host it, and when a mid-sized company should start paying attention.

What exactly did Mistral announce?

In its official announcement, Mistral describes Large 4 as a general-purpose, multimodal model (it reads both text and images) released as "open weight", meaning the model's parameters will be downloadable and runnable on your own hardware. The company says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European datacenters, and that the weights will be released by the end of the month.

Architecturally, it is a Mixture of Experts (MoE) model: the network is split into many specialized sub-networks, and only a small share of them activates for each generated token. Mistral's model documentation lists 1.05 trillion total parameters, 52 billion active per token, a 1.6 billion parameter vision encoder and a context window of one million tokens. Other Mistral pages mention 49 billion active parameters. The gap makes no practical difference, but it is a reminder that this is still a preview.

The launch comes a month after Mistral's €3 billion funding round, announced on September 8, 2026, which values the company at around €21 billion according to Blog du Modérateur. Mistral clearly has the means to keep up the pace, and Large 4 is its strongest bid yet to close the gap with the leading labs.

Is Mistral Large 4 actually competitive?

It pays to separate the vendor's numbers from independent measurements.

According to Mistral's published results, as reported by IT Social, the model scores 61.7% on the DeepSWE v1.1 software engineering benchmark, 59.9% on AutomationBench (agentic tasks) and 93% on Cybench for cybersecurity.

The independent benchmark firm Artificial Analysis paints a more nuanced picture. The preview scores 38 on its Intelligence Index, which by their account makes it the most capable model built outside the US and China, on par with GPT-6 Luna. But the model is very verbose: it generated about 200 million tokens to complete the full test suite, against a median of 81 million for comparable models. As a result, one index task costs $1.13, or $0.57 with the launch discount.

Our take: Large 4 is not the best model on the market, and it does not claim to be. It closes much of the gap with leading American models and sits well above anything a European model has offered so far. For most business use cases (answering questions from internal documents, drafting, classifying, extracting information), that level is more than enough. For long, demanding reasoning, closed models keep the edge.

How much does Mistral Large 4 cost?

There are two ways to access it, with very different cost logic.

Through Mistral's API. The pricing page currently shows $0.68 per million input tokens, $0.07 for cached input and $2.09 for output. Mistral's changelog states this is a 50% launch discount for two weeks, so the reference price is double: roughly $1.36 input and $4.18 output. That second figure is the one to budget with. IT Social notes that even at that price, Large 4 remains cheaper than Claude Sonnet and GPT-5.6 Terra.

Self-hosted. Once the weights are out, cost no longer depends on token volume but on hardware. That is where the model gets demanding (see the next section).

One thing not to overlook is the verbosity measured by Artificial Analysis. A model that produces twice as much text to reach the same result costs more in practice than its per-token price suggests. Before comparing price sheets, always measure cost per real task on your own data. We break down that method in our analysis of what an AI agent really costs.

Can you host Mistral Large 4 yourself?

In theory, yes. For most mid-sized companies, in practice, no.

The catch with MoE models is that only a fraction of the parameters work on each token, yet all of them must sit in memory. The math is simple: 1.05 trillion parameters stored at 4 bits (a common compression technique called quantization) take about 525 GB. At 8 bits, that rises past 1,000 GB. On top of that comes working memory, which grows with the length of the documents being processed.

In concrete terms, even compressed, the model requires a server with eight high-end GPUs of 80 GB or more (640 GB of memory in total), with very little headroom. That is a heavy investment, whether bought or rented, and it only pays off at very high request volumes. For a company handling a few thousand requests a day, the API will almost always be cheaper.

Two questions remain open at the time of writing:

• The license. Mistral has not yet published the terms for the weights. The company's previous models shipped under different licenses, some more permissive than others. Until the text is out, there is no way to know whether commercial use on your own servers will be free or require a contract.
• The date. The announcement says end of month. An announced date is not a delivered date, and a preview can still change before the final release.

For most sovereignty requirements, a smaller open-weight model that runs on a single GPU remains the sensible choice. We compare the options in our guide to open-source LLMs for business.

What does it change for data sovereignty?

For a European company, this is the real story behind the launch.

Until now, organizations under strict constraints (health data, HR data, trade secrets, public procurement) had to choose between a capable closed model hosted outside Europe and an open model they could host but that was noticeably weaker. Large 4 narrows that trade-off: a model designed and trained in Europe, close to frontier level, that you will be able to run on a trusted cloud or on your own premises.

It does not solve everything. Using Mistral's API is still a data transfer to a vendor, with a data processing agreement to sign and processing records to maintain, as with any provider. Self-hosting means securing the infrastructure, logging access and documenting the system. GDPR and the EU AI Act apply either way, whatever the model's nationality. We help companies frame these questions in our GDPR and AI Act compliance work.

If you are weighing public cloud, SecNumCloud-qualified cloud and on-premise servers, our decision matrix still holds with this new model: Large 4 adds a credible option to the right-hand column, it does not change the criteria.

Which use cases is Mistral Large 4 a good fit for?

Three situations justify testing it right away.

Assistants built on internal documents. An agent that answers employees from procedures, contracts or product sheets (known as RAG, for retrieval-augmented generation) benefits directly from the one million token window and the model's comprehension. When those documents are sensitive, the prospect of European hosting carries real weight. We detail this approach on our RAG on internal documents page.

Mixed-format document processing. Scanned invoices, plans, handwritten forms: native image understanding avoids chaining an OCR tool with a language model, and reduces errors at the seams.

Organizations that have ruled out American models. For a public body, a healthcare provider or a company that made that choice as internal policy, Large 4 becomes the most capable option available.

Conversely, for a high-volume customer support agent, a smaller, cheaper model will often do just as well. Large 4's verbosity and reference price make it a model to reserve for tasks that genuinely need it.

Key takeaways for your business

Test on the API, don't migrate yet. The preview lets you measure quality on your own cases while the launch discount runs. Prepare 30 to 50 real questions with their correct answers, and compare Large 4 against the model you already use.

Budget at the reference price and per task. Rely on neither the promotional price nor the headline price per million tokens. What matters is the cost of a complete answer for your use case.

Wait for the license before any hosting project. The choice between API and in-house servers depends on the terms published with the weights. If self-hosting a model this size is out of reach, a more compact open model on a trusted cloud remains the most realistic path.

Frequently asked questions

Is Mistral Large 4 free?

Not through the API: it is billed per token, with a 50% discount during the two-week launch window. Once the weights are published, the model will be downloadable, but commercial use terms will depend on a license that has not been announced yet.

Is Mistral Large 4 more sovereign than ChatGPT or Claude?

The model is designed and trained in Europe, and it will be hostable on the infrastructure of your choice, which is not possible with closed American models. Through Mistral's API, however, your data is still processed by a vendor, with the same contractual obligations as anywhere else.

Can a mid-sized company run Mistral Large 4 on its own servers?

Rarely. Even compressed, it needs a server with eight high-end GPUs, a budget out of reach for most mid-sized companies. A smaller open-weight model or a trusted cloud is a better fit.

Should you replace your current model with Mistral Large 4?

Not without measuring it on your own data first. It is very strong for a European model, but it still trails the best closed models on the most complex tasks, and its verbosity can drive up the bill.