Aleph Alpha Kolibri: German Sovereign LLM Explained
Original: Aleph Alpha Kolibri: How the sovereign German LLM works
Why This Matters
Proves EU-built, regulation-compliant LLMs can compete on benchmark performance at scale.
Aleph Alpha released Kolibri on Oct 3, 2026: a 78B-parameter MoE LLM for German and English, Apache 2.0 licensed, trained on 768 NVIDIA B200 GPUs in Germany and Finland on ~24 trillion tokens.
Kolibri is Aleph Alpha's open-weight large language model built specifically for German and English, using a mixture-of-experts (MoE) architecture with 78.1 billion total parameters—but only 3.46 billion (about 4.4%) are activated per token. Weights are on Hugging Face under the Apache 2.0 license, though Aleph Alpha retains rights to training code and methods.
The model supports a native 262,144-token context window, tested up to 1,048,576 tokens, and has a knowledge cutoff of June 18, 2026. Training used roughly 24 trillion tokens, over a fifth of which were German, run on 768 NVIDIA B200 GPUs hosted in Germany and Finland.
Aleph Alpha defines 'sovereign' along two axes: how it was built (entirely under European and German law, no foreign control) and what customers get (full deployment freedom, IP safety, data never leaving their servers). The company has also signed the EU's General-Purpose AI Code of Practice.
The model card acknowledges outside tooling: English web text was rephrased using Google's Gemma 4, German text with Mistral-NeMo, and Qwen3-32B was used to label data for quality filters. Training data was additionally filtered for political bias identified in Chinese open models. Aleph Alpha claims Kolibri outperforms every comparable model of its size in both German and English on their internal benchmarks.
Features include four reasoning levels (none, low, medium, high), tool calling, and approximately 78 GB of FP8 weights.