Most large language models do their thinking in English, even when you ask them something in another language. The internal chain of thought — the step-by-step scratchpad that has become standard in reasoning models — is overwhelmingly English by default, because the training data, the benchmarks and the researchers' own working language are. Kolibri-1, as Heise reports, is built the other way round: it reasons in German.
It is a deceptively simple design decision, and it reopens an old question. If a model reaches its answer by way of English, the reasoning has to survive a double translation — question into English, answer back out — and every idiom, compound noun and technical or legal term without a clean English equivalent is a chance to lose something. German is a useful stress test: its long compounds and case system carry information that English paraphrases away.
The counter-argument is equally familiar. What matters is the quality of the reasoning, not the language it happens in, and a smaller model reasoning in its 'native' language may simply reason less well than a giant reasoning in English. Heise's framing — thinking in German is one thing, does it help the answers? — is the right way to put it, and the honest answer is that it has to be measured task by task rather than asserted.
The story lands in a European debate that is no longer only about values. Sovereign-AI efforts across the continent have argued for years that models trained and evaluated in European languages behave better on European tasks; sceptics replied that scale decides everything and language is decoration. A model that puts the language of thought itself on the table is a cleaner experiment than most sovereign-AI marketing: it isolates one variable and asks whether it shows up in the output.
Expect the interesting results to be uneven. Translation-heavy work may improve, factual recall in German may sharpen, and English-centric benchmarks may look untouched. What will decide Kolibri-1 is not whether its scratchpad is German, but whether anyone can show the German scratchpad producing an answer that an English one got wrong.
