The French AI company Mistral has presented a new top model with Large 4. According to the provider, it is the most powerful open model from Europe and the USA, which is also intended to reduce the gap to Chinese open-source models. Independent comparative tests that confirm or relativise this classification are still pending – so the performance claim should currently be read with corresponding caution.
The announcement is still interesting, and less because of individual benchmark values than because of the question of what basis companies will build their AI functions on in future.
Why open models are relevant for companies
An open model is an AI model whose weights – i.e. the trained parameters – are published. In principle, it can therefore also be operated outside the manufacturer's cloud, for example on own or rented infrastructure in Europe.
For small and medium-sized companies, this results in three practical advantages:
- Data sovereignty: The processing location and scope of processing can be determined independently. This simplifies data protection impact assessments and the argumentation towards customers whose content could end up in prompts.
- Predictability: Anyone who hosts a model themselves is not affected if a provider discontinues a model version, changes prices or silently adjusts the behaviour of an endpoint.
- Portability: Open weights reduce dependence on a single provider because switching to another operating environment remains technically possible.
This is offset by costs and effort. Large models require powerful GPU hardware, and operation and monitoring are not a side issue. In many cases, managed access via an API – i.e. the provider's programming interface – is more economical, as long as the contractual and data protection framework conditions are right.
European providers as part of the risk assessment
A provider's origin does not replace an assessment, but it changes the starting position. A European company with processing in the EU reduces the justification effort for third-country transfers and makes discussions with data protection officers and works councils easier.
At the same time, the following applies: what remains decisive is the specific contract, the actual processing location and the question of whether inputs are used for training. These points must be clarified for each product and tariff and cannot be derived from the company address.
Coding assistants: the obvious use case
In development teams, AI assistants are now part of everyday life – from code completion and test drafts to explaining unfamiliar codebases. This is precisely where the data question weighs heavily, because source code often contains business logic, access data in configuration examples or customer data in fixtures.
A powerful open model expands the options: teams can operate assistance functions in a controlled environment instead of sending internal repositories to a cloud whose processing chains can only be traced to a limited extent. However, whether the quality is sufficient in your own technology stack can only be clarified empirically – for example by comparing it on real, representative tasks from your own day-to-day project work.
What can be derived from this for the architecture
The model landscape is moving quickly. Anyone making a decision today should make it in such a way that it can be revised in twelve months. An approach in a few steps has proven itself:
- Introduce an abstraction layer: Applications do not communicate directly with a specific model, but with an internal interface. A model change thus becomes a configuration question rather than a rebuild question.
- Build your own test cases: A small collection of typical tasks with expected results shows more reliably than any benchmark whether a model is suitable for your own use case.
- Define data classes: Which content may be sent to external services, which only to your own environment? This classification often decides more about the model choice than performance.
- Weigh costs against benefits: Self-hosting usually only pays off from a certain usage volume or with particularly sensitive data.
Assessment for web and software projects
The announcement of Large 4 is above all a signal that European providers are playing in the upper performance segment and that open models are no longer a stopgap. For ongoing projects, this means: it is worth building AI functions to be interchangeable from the outset instead of binding yourself to a single provider ecosystem. Anyone facing a decision should test models on their own content and clarify data protection requirements before choosing a model – the technology then follows the requirement and not the other way round.