Kolibri shows that sovereign AI can have Chinese teachers

Aleph Alpha has released Kolibri, a German-English Mixture-of-Experts model with 78.1 billion parameters. It activates 3.46 billion of them for each token, or 4.4%. The weights are available under Apache 2.0, and the model supports contexts of up to 1 million tokens.

Kolibri was built in Europe but its post-training drew on China

The company says German teams developed the model and training ran on infrastructure in Germany and Finland. Training covered 24 trillion tokens, with German making up more than 20% of the data. Legal jurisdiction, control over development and the ability to run downloaded weights on a buyer's own infrastructure are the core of its sovereignty claim.

The technical report also names outside helpers. The main models used to generate and regenerate part of the fine-tuning data were GLM-5.2, GLM-5.3 and Qwen3.8-27B. Pre-training also included 4.8 trillion synthetic tokens. Gemma 4 rephrased English documents, Mistral-NeMo rephrased German material, and Qwen3-32B judged a sample of English data.

Sovereignty is about operational control, not a pure pedigree

Kolibri exposes a useful distinction between technological self-sufficiency and operational sovereignty. A European agency or manufacturer can keep the weights, data and inference on its own infrastructure even when foreign models served as teachers and filters during development. For a buyer, the practical questions are who can switch the system off, change its terms or gain access to operational data.

The origin of training influences still matters. A teacher model can pass its errors, preferences and political biases into synthetic answers. Aleph Alpha says it selected models by task and deliberately filtered for those biases. A disclosed trail that lets outsiders inspect those choices is more valuable than the fiction of a European lab working in isolation.

Open weights alone cannot guarantee provenance or independence

Apache 2.0 gives users broad freedom to deploy the weights. The license does not by itself make every dataset, filter and development choice reproducible. Company-run benchmarks also remain company-run until independent tests confirm performance on German documents, tool use and long contexts.

The Chinese-teachers headline is therefore not evidence of deception. It is a reminder that sovereign needs a precise definition. Kolibri offers European jurisdiction and deployment control, not a development chain untouched by foreign models.

Public evaluations will reveal how much control buyers really gain

The next test is independent evaluation and operation outside Aleph Alpha's lab. Watch whether organizations deploy the model without a foreign API, whether they can audit its behavior and how it compares with models activating the same 3 to 6 billion parameters per token.

Kolibri arrives with an unusually detailed account of its data and teachers. If European procurement starts demanding the same transparency from every vendor, the argument over model origin may finally produce something more useful than a geopolitical label.

Lilith's verdict

Kolibri carries a European passport, but GLM and Qwen stood at the front of its classroom. Sovereignty is measured by who controls the switch and opens the class register, not by the teacher's nationality.