Show HN: Germany's new sovereign AI model Kolibri
4 hours ago
- Kolibri is an open-weight large language model from Aleph Alpha for German and English, released on 3 October 2026 under Apache 2.0, with 78.1 billion total parameters but only 3.46 billion active per token.
- It was trained from scratch on infrastructure in Germany and Finland, designed as a sovereign AI that runs fully on customer-controlled hardware under European law.
- The model uses a mixture-of-experts architecture with 384 experts per layer and a router that activates 6 per token, plus a specialized tokenizer (UniBPE) that reduces German token count by 11-15% compared to GPT-5's tokenizer.
- 40 of its 50 layers use sliding-window attention for efficiency, enabling native context up to 262,144 tokens and validated up to 1,048,576 tokens.
- It reasons in German on German prompts, with adjustable reasoning effort (none/low/medium/high) and is trained to abstain from answering when evidence is insufficient (Merlin-Arthur protocol), achieving 44% abstention rate on unknown questions.
- Strengths include top math scores (AIME 2025: 96.9 English, 87.5 German) among models of its active parameter size, strong performance on long-context and company-document tasks, and sovereign deployment.
- Weaknesses include lower factual recall (closed-book test), weaker multi-turn tool calling and coding agent performance, memory requirement of ~78 GB GPU memory, support for only German and English, and reliance on a new vLLM plugin.
- Kolibri is ideal for German-language, sovereign applications like public authorities, banks, and healthcare where data privacy and abstention matter, and for RAG over long German documents; it is not suited for coding agents, trivia, or low-memory environments.