The French artificial intelligence company Mistral presented a preview of Mistral Large 4 on October 6, the largest model it has developed to date. The company affectionately nicknames it "le Chonk" and has confirmed that it will release its weights—meaning the complete model so that anyone can download and run it—by the end of this month.
The announcement comes at a time when Europe is seeking its own alternatives to the large artificial intelligence models developed in the United States and China. The fact that a European company is training and hosting a model of this size in its own data centers has direct implications for companies looking to reduce their technological dependence on foreign powers.
What is Mistral Large 4
The new model has 1 trillion parameters, although it only activates 49 billion each time it generates a response. This design allows for resource savings by not powering the entire system for every query, but only the part necessary to solve the task.
Large 4 can process both text and images and works with more than 160 languages, including all official languages of the European Union. According to Mistral, the model was trained from scratch using 3,800 state-of-the-art NVIDIA graphics cards on its own infrastructure located in Europe. The current preview also runs on those same data centers.
Where to test it and how much it costs
For the moment, Large 4 can be tested in the preview of Mistral Studio, the company's platform, and also via API, the usual route for developers. According to the pricing list published by Mistral, the preview costs $1.36 per million units of text sent to the model and $4.18 per million that the model returns as a response.
The figures presented by Mistral
The company claims that Large 4 ranks among the best open models in the world and that it outperforms any other open model developed in the United States or Europe. According to Mistral, the system competes directly with the most advanced open models globally, including those from China.
In the field of cybersecurity, the company states that its model ranks in the top five in the world according to the independent Artificial Analysis Cyber Index. In a test that consists of reproducing a real software flaw and then repairing it, Mistral claims to achieve 82%, the highest score among all evaluated models. It also claims to solve 93% of the challenges in Cybench, a collection of 40 exercises from computer security competitions.
In programming, the company reports 61.7% in the DeepSWE v1.1 test and 28.3% in Terminal-Bench 4.
A detail that Mistral highlights especially is that, in the aforementioned cybersecurity test, other closed models like Claude Opus 5.5 and GPT-6 Astra obtain results close to zero because they refuse to perform the task. The company argues that defense teams need an artificial intelligence without those types of restrictions to be able to work effectively.
While the release of the weights is being finalized, Mistral states that it is testing the model with cybersecurity managers, selected partners, and state authorities, applying less moderation and more capabilities oriented toward this field.
What this means for companies and users
For European companies, having an open model of this level, hosted on the continent's own servers and capable of being installed on private infrastructure, represents a real alternative to depending on the major American or Chinese tech companies.
Users who are not involved in programming will not notice immediate changes. The impact will arrive indirectly, through the tools and services that integrate this model into their internal operations.
Regarding whether it is free, although Mistral has confirmed that the weights will be downloadable by the end of the month, the company has not yet detailed under what license they will be published. The fact that a model can be downloaded does not necessarily imply that its commercial use is free, so it will be necessary to review the exact terms when they are published.
It is worth remembering that all the performance figures mentioned come from Mistral itself. Until independent third parties can test the model once its weights are released, these data should be treated as company statements and not as externally verified results.
Frequently Asked Questions
What is Mistral Large 4?
It is the largest artificial intelligence model developed to date by the French company Mistral, with 1 trillion parameters, capable of processing text and images in more than 160 languages.
When will I be able to download Mistral Large 4?
Mistral has announced that it will release the model's weights, meaning the downloadable version to install on your own servers, by the end of October 2026.
Will Mistral Large 4 be free?
It will be downloadable, but the company has not yet specified under what license it will be published. The fact that a model can be downloaded does not automatically guarantee that its commercial use is free.
Where can I test the model now?
Currently, the preview is available in Mistral Studio, the company's platform, and also via API access for developers.
Is it true that Mistral Large 4 is better at cybersecurity than other models?
According to data provided by Mistral itself, the model obtains high results in tests such as Cybench and in software flaw repair. However, these figures have not yet been verified by independent third parties.
Where was this artificial intelligence model trained?
Mistral states that Large 4 was trained from scratch in its own data centers located in Europe, using 3,800 state-of-the-art NVIDIA graphics cards.