Germany’s sovereign AI model Kolibri challenges US and Chinese rivals
Built in Germany from scratch.
Image by Insaanu Studio via Unsplash/Illustration by Cybernews
Follow us
- Aleph Alpha released Kolibri, a bilingual AI model built in Germany for regulated organizations.
- Kolibri has 78.1 billion total parameters and can run on-premises to keep internal data off third-party services.
- Aleph Alpha says Kolibri beat Alibaba, Mistral, and Nvidia models in reasoning, coding, and information-navigation tests.
- Kolibri still trails Alibaba’s Qwen model in long-context comprehension and retrieval tasks.
Key Takeaways by nexos.ai, reviewed by Cybernews staff.
The new open-weight large language model (LLM), designed for public administration, is outperforming competitors from Alibaba and Nvidia.
The Heidelberg-based AI startup Aleph Alpha released Kolibri on October 3rd, 2026, which is German Unity Day.
Kolibri is an English-German Mixture-of-Experts Transformer, meaning it is divided into separate sub-networks called “experts.” The model uses 3.46 billion active parameters per token, out of 78.1B total.
The model was specifically built for “sovereign mission-critical work” in regulated areas, including public administration, industrials, and aerospace, taking into account their particular “domain-specific language, regulatory, and procedural realities.”
The company says Kolibri’s small, efficient size allows it to run on-premises without sending internal data to third-party inference services.
Kolibri complies with the EU AI Act, the General-Purpose AI Code of Practice, and the General Data Protection Regulation (GDPR), according to Aleph Alpha, with “copyright law being a focus of our work to trustworthy technology.”
The model was built in Germany and trained on infrastructure in Germany and Finland, under the law of these countries.
It was specifically trained to answer “I don’t know” when it cannot provide an answer, reducing the risk of hallucinations.
Kolibri was developed as a German and English bilingual model, including organic German data – 21.3% of the pre-training tokens are German.
“We started the year with the conviction that Europe needs its own sovereign AI. Built in Germany from scratch and on infrastructure in Europe specifically for our customers,” Aleph Alpha’s X post reads.
Kolibri vs American and Chinese models
Aleph Alpha compared Kolibri to 3 other open-weight models with up to 4 times its active parameter count: Alibaba’s Qwen 3.6-35B-A3B, Mistral Small 4 119B-A6B, and Nvidia’s Nemotron 3 Super 120B-A12B.
Kolibri outperformed all rivals across mathematical reasoning in both German and English (AIME 2025 and 2026 benchmark), deep reasoning (GPQA diamond), and code (LiveCodeBench v6).
The model also showed superiority across benchmarks assessing an AI agent’s ability to find and navigate through large amounts of information.
Kolibri, however, fell behind Qwen’s 3.6-35B-A3B across long-context comprehension and retrieval.
Where to download Kolibri?
The model can be downloaded from Hugging Face and used under the Apache 2.0 license terms.
It requires the Aleph-Alpha-inference package, which provides the Kolibri vLLM plugin; the plugin can be installed from here.
What are other European alternatives?
Kolibri is released as Europeans look for alternatives to US tech, including products from leading AI labs.
In July, a German research consortium, Soofi, released an open-source LLM called Soofi S that outperforms larger counterparts in English and German benchmarks, and shows strong performance in coding, maths, and regional knowledge of German.
Mimir, a 1B-parameter model developed by the University of Southern Denmark and the AI company Ordbogen A/S, has also demonstrated capabilities comparable to those of larger models in the 2-3B and 4-5B ranges, including Gemma 4 E2B.
