Mistral Large 4 is now open weight: a trillion-parameter model you can run in Europe

Mistral, the European AI company from Paris, has released the weights of its flagship model Mistral Large 4, nicknamed Le Chonk. The model has around one trillion total parameters, about 50 billion of them active per token, and it is a natively multimodal mixture-of-experts model with a one million token context window, trained from scratch on European infrastructure. For a European business, this is the most important open model release of the year, because it offers a real alternative to the closed models from the United States.

What open weight means

Open weight means you can download the model's weights and run it on your own infrastructure, adapt it and fine tune it. You do not need to ask Mistral for permission, and your requests do not go to a server in the United States. That is not the same as fully open, because the training material and the training data stay with Mistral. What is open is what you need to run it yourself.

The licence is the Mistral Ecosystem Open-Weight licence, a modified MIT licence. It allows free use, modification, fine tuning and commercial use. There is one condition that matters: if you offer the model as a service to third parties, as model-as-a-service, you need a separate commercial agreement with Mistral. If you run it for your own staff or build it into your own product, that condition does not apply.

What the model can do

Mistral positions Large 4 as a model for agents and for cybersecurity. According to the company's own numbers, it reaches 41 percent on Terminal-Bench 4.0, a jump of 13 points from the public preview, and 48 percent on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47 percent fewer tokens per task. On refusing harmful requests, Mistral reports 96.6 percent refusal, the highest among open weight models, and 97.1 percent attack resistance on the B3 Agent Security Benchmark.

These numbers come from the maker, and they should be read that way. The 41 percent score was measured in Mistral's own agent system Vibe, not in a neutral agent. Wait for independent testing before you move a production system. The direction is still right: an open model that can write code, find vulnerabilities and check patches is something new for a European team that wants to keep its data.

The model works with text and images, knows more than 160 languages including all official EU languages, and offers three reasoning modes: low, medium and high, with high as the default. API pricing is 0.70 US dollars per million input tokens, 0.07 dollars for cached input, and 2.10 dollars for output.

What changes for a European business

First, data residency. You can run a model at frontier level in your own environment, in a European data centre, under your own rules. Before this release, that was not possible with an open model.

Second, cost control. If you run a coding agent or a security check often, you pay per token at a provider. If you run the model yourself, you pay for hardware and power, and that can be cheaper at high volume, even though the start is more expensive.

Third, control over behaviour. You can fine tune the model with your own data, for example your product documents, your policies or your customer processes, without that data travelling to a provider in another country.

The honest limit: self hosting is not free

A model with one trillion parameters needs serious hardware. A small business will not run it on an office laptop. The practical routes are: use Mistral's API, run the model through a European hoster, or use a smaller fine tuned version for one specific task. The weights are available on Hugging Face, and they run with vLLM or SGLang, or as an NVIDIA NIM container.

Three steps to take this week

First, decide whether you need your own model at all. For most businesses, the API is the right route, because it needs no hardware and the costs are visible.

Second, if you work with customer data or sensitive documents, check whether a European hoster or your own instance of Large 4 is cheaper than a US provider once you count the data residency rules.

Third, test the model on a real task before you switch. For example, let it build an agent that checks your code base, or summarise a security log. Measure the quality and the cost over one week, then decide with numbers.

The takeaway

Mistral Large 4 gives European businesses something that did not exist before: an open frontier model, trained in Europe, that they can run themselves. It does not solve every problem, and the maker's own benchmarks are no substitute for your own tests. But for a business that wants to keep its data in Europe and runs a coding agent or a security check, it is from today the first choice among open models. Read our guide to sovereign AI for SMEs to see how your own model fits into your digital systems.