Aleph Alpha Kolibri: How the sovereign German LLM works
What happened
Aleph Alpha released Kolibri, an open‑weight large language model for German and English on 3 October 2026 under Apache 2.0, with 78 billion parameters but only about 3.5 billion active per token thanks to a mixture‑of‑experts architecture. The model was trained from scratch on infrastructure in Germany and Finland, aiming for compliance with the EU AI Act and offering full deployment freedom and IP safety for customers. Kolibri uses 50 layers, each with 384 experts plus a shared expert, routing each token to six experts, which reduces compute while keeping performance high. The team highlights that the model’s training data included re‑phrased English web text from Google’s Gemma 4, German text from Mistral‑NeMo, and quality‑filter data from Qwen3‑32B. The launch includes a 189‑page technical report, model card, and a tokenizer experiment from the author.
Why it matters
For German businesses and public institutions, Kolibri provides a locally hosted AI that respects data privacy and regulatory constraints, potentially reducing reliance on foreign services. It also shows how mixture‑of‑experts can deliver large‑scale language capabilities with lower per‑token compute.
Reality check
An automated, sceptical read of the sourceIncremental. The model introduces a clever MoE efficiency trick and sovereign deployment, but builds on existing expert‑routing concepts rather than presenting a fundamentally new breakthrough.
Is this a big deal?
Get stories like this every morning
The day's AI news in plain English, in your inbox at 8:30 IST. Free.
More AI news
AI unlocks Gödel's notebooks, said to contain hundreds of proved theoremsIndustry news · 13 Nov 2026Gemini 4 Hydrogen and OrchestratorIndustry news · 12 Nov 2026Gemini 4 HydrogenIndustry news · 12 Nov 2026I'm making LLMs work on small and old devicesIndustry news · 12 Nov 2026United Airlines bans AI agents from booking flights on behalf of travelersIndustry news · 11 Nov 2026This page was collected from tej.as and summarised automatically by AIPulsee's pipeline. See how it runs.