Aleph Alpha Kolibri: How the sovereign German LLM works
Aleph Alpha released Kolibri, a 78-billion parameter mixture-of-experts open-weight LLM with 3.5 billion active parameters per token, optimized for German, long context, and regulatory compliance.
read it The post offers a rare, detailed breakdown of a European open-weight LLM's architecture, tokenizer, and training mechanics.
summary
Aleph Alpha's Kolibri-1 is a 78-billion parameter mixture-of-experts model featuring 3.5 billion active parameters per token and a 1-million-token context window designed around EU AI regulations. Built with a custom UniBPE tokenizer tailored for German compound words and a training protocol discouraging hallucinations, it achieves competitive scores in German math and document QA. However, it requires significant GPU memory (78 GB) and lags behind dense models in general memory tasks and coding agents. The article details its specifications, performance tradeoffs, and practical deployment steps using vLLM.
key points
- Kolibri uses 50 layers with 384 experts each, routing each token to 6 experts to achieve 3.46 billion active parameters out of 78.1 billion total.
- Its UniBPE tokenizer requires 11.2% fewer tokens for German text than GPT-5's tokenizer by natively handling compound words.
- The Merlin-Arthur training protocol forces the model to abstain and state I do not know when evidence is missing, succeeding 44% of the time on unknown queries.
- It demands substantial resources, requiring at least 78 GB of GPU memory to hold the full weights in memory.
who should read it
AI engineers and developers looking for sovereign, open-weight LLMs with strong German-language capabilities and long-context RAG support.
discussion mixed
consensus
Commenters largely focus on Hacker News submission etiquette rather than the model itself, noting that the post was incorrectly tagged as a Show HN since the blogger did not build the model.
pushback
None regarding the model's technical merits, though some debate the appropriateness of the submission format.
The thing to note here, besides the transparency and the fact that it’s actually a good model that also works well on coding and agentic tasks, is that it’s the first release by a team formed less than a year ago, with a strong focus on iteration velocity. There’s more to come.
disclaimer: I‘m part of the training team, happy to answer any questions
A training team member notes the model is the first release by a newly formed team with rapid iteration velocity, offering brief insider context.
Generated by an LLM from the article and thread. Check the sources before relying on it.