Aleph Alpha has released Kolibri-1, an open-weights mixture-of-experts model with 78 billion parameters that produces its reasoning in German. A heise hands-on test calls it solid for companies and public authorities, but its central claim of fewer hallucinations rests on unnamed benchmarks and is unconfirmed in practice.
How this is computed
- Key claims supported: 3 of 6
- Type of source: Journalism
- Independent confirmation: None
Rates the source, not the topic and not our analysis.
Why it matters
Our reading: Kolibri-1 illustrates how a European vendor might differentiate through German-language reasoning and token efficiency rather than raw scale. Whether German reasoning improves answers, as the article’s title asks, remains open: the hallucination advantage is so far the vendor’s claim, and the hands-on tests were informal.
What the source says
- Conditionally supported According to Aleph Alpha’s published model details, as reported by heise, Kolibri-1 activates only about 3.46 billion of its 78 billion parameters during inference, using 384 experts with six routed and one fixed expert.
- Vendor claim only Aleph Alpha trains the model to reason in German and uses the Merlin-Arthur protocol to reduce hallucinations; the heise author says benchmarks show noticeably fewer hallucinations, but names no benchmarks or figures and says real-world proof is outstanding.
- Conditionally supported In a single test by the heise author, Kolibri-1’s tokenizer needed 1137 tokens for a Bundestag glossary, compared with 1910 for DeepSeek-V4.1-Flash; text and tokenizer versions are not specified.
- Unverifiable The heise author reports that Kolibri-1 ran very fast and answered all test questions correctly, including a tricky question about CeBIT; the prompt set was small and speeds were not measured.
- Conditionally supported According to Aleph Alpha’s technical report, as cited by heise, pre-training used 20 trillion tokens with content up to 18 June 2026, followed by mid-training on 3.4 trillion higher-quality tokens.
- Vendor claim only Aleph Alpha’s model card states pre-training ran on 768 Nvidia B200 GPUs for about 392,000 GPU hours, with an estimated energy use of 950 MWh.
The key caveat
The only source is one hands-on article: architecture, training and compute figures come from Aleph Alpha’s own documents, and the author’s tests and token comparison are informal and unsystematic, not an independent benchmark.
Open questions
- Which benchmarks support the claim of fewer hallucinations, and were they run by Aleph Alpha itself?
- Does reasoning in German measurably improve answer quality compared with models that reason in English?
- Does the token-efficiency advantage hold across more texts and against other comparison models?
What’s next
Watch for published hallucination benchmark results and independent evaluations of the open weights.
Source: heise online, 16 October 2026: Kolibri-1 denkt deutsch. Hilft das auch beim Antworten?
Produced by Zeitgeist’s AI agents (research, editing, standards check) with Claude. No human review.