Mistral Large 4 makes the case for sovereign weights

Mistral's preview of Mistral Large 4 pairs a 1-trillion-parameter mixture-of-experts architecture (activating 49 billion parameters) with an explicit push for infrastructure sovereignty. Training a model of this scale on 3,800 Grace Blackwell GPUs inside European data centers matters because open-weight parity at the high end has largely been dominated by Chinese labs over the past year. Having an EU-based lab deliver frontier-competitive weights gives European organizations a viable alternative to hosted US APIs without sacrificing capability.

The most instructive benchmark in their release data is cybersecurity evaluation. On tests requiring a model to reproduce and patch real software flaws, Mistral reports ML4 scores 82 percent while closed frontier models often score near zero because platform safety filters classify vulnerability reproduction as malicious. When provider-level refusals shut down legitimate penetration testing or incident analysis, enterprise security teams cannot depend on external black-box endpoints. Self-hosting an open-weight model on private infrastructure removes that friction, turning model weight ownership into a practical operational requirement rather than an ideological stance.

While the benchmarks look promising, I am looking forward to putting it through its paces, so I can see for myself how it performs.