Mistral goes Apache again — and 'open source' stops meaning 'second best'

A European lab moving its flagships back to a permissive licence is a small event with a large meaning: the open-weight frontier now has terms, and they are the terms the market settled on.

Mistral relicensed both Large 3 and Small 4 under Apache 2.0, reversing its earlier drift toward restrictive terms on its larger models. The licence is the substance, not the model number: Apache 2.0 permits commercial use, modification, and redistribution with minimal obligation, which is the difference between a model an enterprise can ship a product on and one it can only evaluate.

Why a European lab moved

The pressure was permissive and largely Chinese. Open weights have stopped merely competing and started winning on coding, math, and long-context tasks, led by Qwen 3 under Apache 2.0, with DeepSeek under MIT. A European lab holding its best models under restrictive terms was ceding the deployable-open-weights market to labs that set the expectation that frontier-adjacent weights come with liberal licences. Relicensing is Mistral choosing to compete on the terms the market has already fixed.

'Winning, not competing' is a real threshold

For years the open story was 'nearly as good, much cheaper.' The claim now is that on specific high-value axes the best open models top the charts outright, which changes the default from fallback to first choice for those workloads. And a top-of-benchmark model that is also freely deployable compounds: it seeds an ecosystem of derivatives, and every project built on it is a reason the next one starts there too.

The compliance angle nobody priced in

There is a timing irony worth noting. As the EU's training-data-summary duties become enforceable, licence terms and provenance become compliance inputs, not just engineering preferences. A Western enterprise wary of the provenance of the leading Chinese open models now has a European Apache-2.0 flagship option — which matters more, not less, in a quarter when a regulator can ask where a model's training data came from. Open weights are becoming normal infrastructure, and their licences are becoming part of the paperwork.

Thunder Compute — Best open source LLMs (August 2026) → · Hugging Face — Best open-source LLM models in 2026: coding, local, agentic, benchmarks, license →