Geoffrey Vandiest
Mistral announced Large 4 this month, with open weights promised for the end of October. Good news, and I mean that. But let's be honest: it still trails the American frontier, and one company is not a strategy. Europe would need five Mistrals, and we have one.
Here is the part nobody in Brussels seems to say out loud. The money in AI is not in building models. China has settled that question by giving its models away. DeepSeek, Qwen, Kimi: open weights, permissive licences, free to download. Chinese models now carry close to 60 percent of the tokens American companies route through OpenRouter. Not because they are the best. Because they are nearly as good and a fraction of the price.
They are also a fraction of the energy. These models are built sparse and small on purpose, because China has less compute to burn. A model that runs on one modest server is a model you can run in your own data centre, behind your own firewall, with nothing leaving the building. Nothing prohibits that. I run open models on my own machine when privacy matters. It is not exotic. It is a download.
So where is the value? In the use. In the hospital that reads its own records with a model it hosts itself. In the public institution that automates a process it has been doing by hand for twenty years. In the three-person startup that builds a product on a free model and keeps its margin. That is where European growth sits, and it is sitting there for free.
Europe should stop mourning the frontier and start shopping. Embrace the free models, run them in isolation, use them creatively. It is perfectly possible to do this safely. The question is not whether we can build the engine. It is whether we are smart enough to drive it.
First published on LinkedIn.