Aleph Alpha releases open weights of 78B-parameter Kolibri model

On 3 October 2026, Germany's Aleph Alpha published the weights of Kolibri on Hugging Face under the Apache 2.0 licence. The mixture-of-experts model has 78.1 billion total parameters (3.46 billion active per token), supports English and German with a 262,144-token context, and was trained on 20 trillion tokens using 768 Nvidia B200 GPUs.

• 78.1B total parameters, 3.46B active per token, 50 layers and 384 experts
• FP8 weights take about 78 GB and need at least 2x A100 80 GB or 1x H200
• Native 262,144-token context, validated up to 1,048,576 tokens
• Trained on 20 trillion tokens on 768 B200 GPUs over 21 days, about 392,000 GPU-hours