Why would anyone build a model that trails?

Here is the awkward bit. A French lab, Mistral, has just released Large 4, nicknamed Le Chonk. A trillion parameters, serious hardware behind it, a proper launch. And on the independent scoreboards it sits behind the strongest open models from China, never mind the big American ones. So why spend a fortune building something that is not number one?

Fine, but if it is not the best, why would anyone pick it over the ones that are?

Because for a certain kind of customer, "best" is not the question. Control is.

Think about a bank, a hospital, a defence contractor, a government agency. They sit on data that cannot casually leave the building. Now add a US law called the CLOUD Act, which lets American authorities demand data from American companies even when that data lives on a server in Europe. For a European ministry, that is a problem no benchmark score can fix.

Open weights are the way out. You download the model, run it on your own machines, in your own country, under your own laws. Nothing leaves. And because you hold the whole thing, you can inspect it, audit it, and check what it is doing.

It is industrial policy too

Same logic as chips, same logic as satellites. Nobody wants to depend entirely on Washington or Beijing for a technology this foundational. A second-tier model you fully control can be worth more than a first-tier model you only rent. And the commercial pitch writes itself: organisations can adapt the model for their own sovereign deployments instead of begging a foreign provider for a special arrangement.

There is a catch, of course. Sovereignty only counts if the model is good enough that people actually want to use it. A sovereign model nobody can work with is just an expensive monument.