Mistral has released the weights of its Large 4 model, which runs internally under the name "Le Chonk". It is a model with one trillion parameters that, according to the provider, was trained entirely in Europe. Open weights means: the trained model parameters are available for download and can be run on your own or rented infrastructure, instead of exclusively via a third-party programming interface.

For companies that want to use AI tools in software development, that is the real news. In practice, it is not the benchmark ranking that decides, but the question of where model weights are located, who has access to the processed data and how predictable the costs are.

What the measurements show – and what they do not

In programming, Mistral sees the model ahead of the open competition. In independent tests, it also stands out for its high speed and comparatively low operating costs. Against the current Chinese top, however, it is not enough in the overall comparison: 26 points there stand against 32 points for DeepSeek V4.1 Flash and GLM-5.3.

These figures are an overall value across different task types. They say little about how a model behaves in a concrete, narrowly defined workflow – for example when generating test cases, refactoring existing classes or migrating templates. A model that lags behind on average can still be the better choice in a specific use case if it is specifically strong there.

Why open weights are relevant for SMEs

The practical difference between a closed API model and an open model shows up in three points:

  • Data flow: Anyone who runs the model themselves or with a host in the EU controls which source code and documents leave their own network. That considerably simplifies coordination with data protection and works councils.
  • Availability: With open weights, a model version that has once been validated can continue to be operated. Closed services change model versions without customers being able to control that.
  • Cost structure: Instead of variable billing per request, there are more likely to be calculable infrastructure costs – relevant as soon as tools are permanently built into development processes.

On the other hand, there are efforts that should not be underestimated: a trillion-parameter model places demands on hardware and operations. For many SME scenarios, therefore, the goal is less self-operation in one's own data centre than hosting with a European provider with clear contractual commitments.

Origin is an argument, but not a free pass

The fact that Large 4 was trained in Europe is interesting for organisations that want to reduce their dependence on US providers. However, that does not replace their own review. What remains decisive is where the model actually runs later, which contract stands behind it and what log data is generated. A European-trained model on a non-European platform does not solve the original problem.

Likewise, open availability does not exempt anyone from the question of licence terms for commercial use. This should be clarified before building a toolchain, not afterwards.

A pragmatic approach

Anyone who wants to examine deployment will usually get further with a lean process than with a broad evaluation:

  1. Define two to three concrete tasks from your own day-to-day work – for example code review suggestions, unit test generation or documenting existing modules.
  2. For these tasks, put together a small but realistic test set from your own project.
  3. Have the open model compete against the service currently in use and have the results assessed by developers, not by a score.
  4. Measure operating costs and response times in the same run, because both influence acceptance in the team more strongly than quality differences in the per mille range.

This form of assessment is more time-consuming than a glance at a ranking list, but it leads to robust decisions. It also has the advantage that the test setup can be reused for future model changes – and model changes are currently more the rule than the exception.

Classification for web and software projects

For projects with data protection requirements, the selection shifts: it is no longer solely the top quality that counts, but the combination of sufficient code quality, calculable costs and a controllable operating location. Large 4 is a serious option for this, even if it does not lead in the overall comparison. It makes sense to integrate AI tools in development in such a way that the underlying model remains interchangeable – then the next model generation becomes a configuration question instead of a project.

Sources