Aleph Alpha launches Kolibri, a German 78.1B-parameter MoE model
Aleph Alpha released Kolibri on October 3, 2026, an open-weights German and English model with 78.1 billion total parameters that activates only 3.46 billion (4.4%) per token. It was trained on 24 trillion tokens using 768 Nvidia B200 GPUs in Germany and Finland, with weights under Apache 2.0.
- 78.1B total parameters, only 3.46B active per token (4.4%)
- Trained on 24 trillion tokens with 768 Nvidia B200 GPUs in Germany and Finland
- Open weights under Apache 2.0, 189-page technical report
- UniBPE tokenizer cuts German token usage by 11.2% versus GPT-5
Read next
Business