A 675-billion-parameter open model, and a gap measured in months

Europe just shipped a frontier-scale open-weight model under a real permissive license, while the leading open labs say they trail the closed frontier by only months. The open option is no longer a compromise.

Mistral Large 3 shipped as a 675-billion-parameter mixture-of-experts model under a true Apache 2.0 license. The mixture-of-experts design keeps only 41 billion parameters active per token, so it delivers frontier-scale capacity while staying deployable — and the permissive license means enterprises can build on it, not just evaluate it.

The gap is now months, not generations

And it is closing across the board. DeepSeek estimates its open weights now trail the closed frontier by only a few months on many benchmarks. When the lag is that short, the calculus flips: for most workloads, an open model months behind at a fraction of the cost is the rational default, not the fallback.

Provenance becomes a feature

A Western, Apache-licensed frontier open model is also a provenance argument. As the EU training-data-summary duties become enforceable and enterprises weigh the risk posture of the leading Chinese open weights, a clearly-licensed European flagship is deployable in ways that matter for compliance, not just capability.

The open field now has a Western flagship at the top tier and a gap to the frontier measured in months. That combination is what turns open weights from the cheap alternative into the default choice.

TechJack Solutions — Best open-source AI models 2026: 5 that rival frontier → · LLM-Stats — AI updates today (August 2026) →