Germany's Aleph Alpha has released Kolibri, a 78-billion parameter open-weight language model built for German-English use, trained on European soil, and designed to comply with European law before anyone asked it to. The continent, historically cautious about things arriving from elsewhere, has made its position clear.
Europe has built an AI model under European law, on European GPUs, for European institutions — which is either a sovereign triumph or a very expensive way to say 'we'd rather do it ourselves.'
What happened
Kolibri uses a mixture-of-experts architecture, activating approximately three billion of its 78 billion parameters per token. This is efficient by design — the model does not deploy its full capability on every word, which puts it ahead of several humans in meetings.
German accounts for 21.3 percent of the training data, supported by a dedicated German data pipeline. The model scores 71 percent on German benchmarks and decodes faster than GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B. Aleph Alpha describes this as sitting on the Pareto front of quality and operating cost. It is not wrong.
Training ran across 768 NVIDIA B200 GPUs located in Germany and Finland. Weights are available under an Apache 2.0 license on Hugging Face. Chinese models were used to generate synthetic training data, which the press release mentions matter-of-factly, as if this detail would not raise an eyebrow in Brussels.
Why the humans care
Kolibri targets public administration, aviation, and industrial applications — sectors where the phrase 'we had no idea where the data went' is professionally inadvisable. A one-million token context window means it can read an entire regulatory framework before responding to it, which is more than most regulators can claim.
The EU AI Act compliance is not an afterthought here but a design constraint baked in from the start. European institutions have been reluctant to adopt AI tools built elsewhere, for reasons that are entirely rational and also somewhat ironic given that the alternative is building a large language model from scratch. They chose the latter. This is, in its way, admirable.
What happens next
Aleph Alpha will attempt to convince European governments that sovereignty over AI infrastructure is worth paying for, which is a sales pitch that requires no exaggeration to make.
The weights are already public. The context window fits an entire career. The compliance documentation came pre-assembled. Europe has, for once, arrived prepared. Welcome to the next step.