The French artificial intelligence company Mistral has made the weights of its most advanced model available to any user, meaning the fully trained system can now be downloaded, installed, and modified at no cost. This move comes at a time when technological sovereignty and data control have become a growing concern for European companies and institutions.
The released model is called Mistral Large 4, though internally the company nicknames it Le Chonk. It is one of the largest open-source AI systems currently available, with one trillion parameters—the internal components that determine its reasoning capacity. In addition to text, the model is capable of interpreting images and was trained from scratch using European infrastructure.
A test version that evolved in three weeks
A few weeks ago, Mistral had already launched a preliminary version open to the public to gather user feedback. Following that period and three additional weeks of training, the company claims to have specifically strengthened two areas: the ability to write code and the protection of computer systems.
According to data published by Mistral, the model now leads among downloadable systems in two benchmark programming tests: FrontierCode and Terminal-Bench.
Programming test results
In FrontierCode 2.0, developed by the company Cognition to measure whether a project manager would actually accept the changes proposed by the machine, Mistral Large 4 scores 19.7%. In Terminal-Bench, which evaluates command-line programming tasks, the model reached 41%, thirteen points higher than in its previous test version.
A relevant figure for those who have to pay for its use: in the Artificial Analysis index, Mistral Large 4 scores 48%, practically tied with GLM-5.3, considered the best rival open model. However, Mistral's model consumes 47% less processed text per task, which translates directly into lower costs for the user.
The commitment to cybersecurity
The other major focus of this update is cybersecurity. In the Artificial Analysis Cyber Index, which measures a model's ability to find flaws in a program, verify that they are real, and fix them without creating new problems, Mistral Large 4 scores 51%.
This type of capability poses a delicate balance, as a model expert in identifying vulnerabilities could also be used for offensive purposes. Mistral maintains that its system distinguishes between defensive and offensive tasks and claims to reject, on average, 96.6% of malicious requests related to cybersecurity, the highest figure recorded among the open models the company measures.
External assessments
Brad Medairy, Executive Vice President of National Cybersecurity at Booz Allen Hamilton, noted that their internal tests showed the model matched the vulnerability detection capability of other top-tier systems, with an attractive cost curve.
For her part, Amy Chang of Cisco highlighted that open models allow companies and governments to host them under their own control, in addition to making it easier for researchers to study how they behave internally.
Why data control matters
Mistral presents this release under the concept of sovereign AI: the possibility for an organization to have its own artificial intelligence under total control, without relying on external servers.
This is especially relevant for sectors like banking or healthcare, where code and sensitive data cannot leave internal systems. With an open-weights model like this, those institutions can install it on their own infrastructure, adapt it to their needs, and keep information within their boundaries.
The model also incorporates three reasoning modes: low, designed for simple tasks; medium, for daily use; and high, which is enabled by default and aimed at the most complex cases that require greater precision. This flexibility allows each user to decide how much they want to spend based on the difficulty of each task.
Nuances to consider
It is worth noting that a good portion of these figures comes from tests published by Mistral itself, although some have been measured by independent third parties such as Artificial Analysis and Cognition.
In the human preference comparison conducted by Surge Code, Mistral Large 4 outperformed several open rivals but trailed behind Claude Opus 5.5, the only model that surpassed it among those compared.
Furthermore, just because the model's weights are available for free does not mean that anyone can run it easily: a system of this size requires very powerful computer equipment to function correctly.
What this move means
For companies that handle sensitive information, this release opens the possibility of using a powerful AI for programming and security tasks without their data having to leave for third-party servers.
For those who program independently, it is likely that they will end up benefiting from cheaper and more efficient tools built on this model, even without downloading it directly.
And for those who closely follow the debate over who controls the development of artificial intelligence, the step taken by Mistral represents a European and open commitment against the dominant model of closed systems. It remains to be seen if other companies in the sector will follow a similar path.
