Aleph Alpha has released Kolibri, an open-weight large language model designed with European regulations in mind. This German-English model boasts 78.1 billion parameters but efficiently uses only about 3.46 billion per token, making it a mixture-of-experts model. Kolibri was trained from scratch on infrastructure in Germany and Finland, emphasizing data privacy and intellectual property safety for deployment within organizations. While trained on European data and adhering to the EU AI Act, it incorporated English web text rephrased with Google's Gemma 4 and German text labeled with Mistral-NeMo and Qwen3-32B for quality filtering. AI
IMPACT Sets a precedent for regulatory-compliant AI development in Europe, potentially influencing future model releases and data governance.
RANK_REASON Open-weight model release from a European AI lab with a focus on regulatory compliance.
Read on Mastodon — fosstodon.org →
- Aleph Alpha
- Germany
- Kolibri
- EU AI Act
- European Union
- Finland
- Gemma 4
- Hugging Face
- Michael Hofmann
- Mistral Nemo
- Qwen3 32B
AI-generated summary · Google Gemini · from 14 sources. How we write summaries →