The sovereignty trade: what picking Mistral actually buys you
Twice this year I've sat in architecture reviews where the model decision took five minutes, because compliance had already made it. Healthcare data that can't leave the EU. A public-sector contract requiring on-prem inference with auditable supply chains. In both rooms, the shortlist was Mistral or nothing — and I think engineers outside these industries underestimate how common these rooms are becoming.
What sovereignty concretely means
Strip the geopolitics and it's three checkboxes. Residency: inference happens on EU soil, with an EU legal entity holding the contract, so there's no adequacy-decision roulette and no transfer-impact assessment every time the political weather changes. Possession: for the strictest cases, weights you can download (Small's Apache-2.0 release, Devstral, the older open lines) running on hardware you physically control, the same stack I described in the sovereign coding setup. Legibility: a vendor whose regulatory home is the EU AI Act itself, which makes your compliance documentation an import rather than a translation.
None of this makes a model smarter. All of it decides whether the project exists.
In regulated Europe, the first model filter is jurisdiction, and most of the leaderboard doesn't survive it.
What it costs
Honesty section. The capability tax is real: on my benchmark, Mistral's top entries trail the frontier tier, and the newest ones haven't published SWE-bench runs at all. I've complained about that, and I won't stop. For agentic coding under a sovereignty constraint, you're choosing between a capable-but-not-peak hosted model and Devstral on your own GPUs, and either way you're giving up the top of the market.
The ecosystem tax compounds it. US-first tooling means EU-first deployments do more integration work themselves: fewer turnkey harnesses, and evals you largely build in-house. Budget for it or it becomes schedule slip with a French accent.
And a warning from a project I watched go sideways: sovereignty theater. An EU model behind a US-owned inference platform with US-controlled telemetry satisfies nobody's regulator. If jurisdiction is the requirement, it's the whole stack: model, serving, logging, support path. Otherwise you've bought a sticker, not a property.
How I'd decide
If no regulation binds you, decide on capability and price like everyone else; sovereignty you don't need is just a smaller model. If regulation binds you softly (data residency but hosted is fine), Mistral's EU API is the path of least resistance and Large 3's price makes it painless. If it binds you hard, open weights on your own metal is the only honest answer, and the interesting news of 2026 is that this option stopped being a toy. The gap between "compliant" and "good" closed enough that choosing both is no longer a contradiction, most days.