Aleph Alpha releases open-weight Kolibri AI model
The German and English model uses 78 billion parameters, with about 3.5 billion active per token. Aleph Alpha says it can handle context windows of up to one million tokens.
Aleph Alpha released Kolibri on 3 October under the Apache 2.0 licence, making the German and English language model’s weights available for use and modification. The company says it was trained from scratch on infrastructure in Germany and Finland.
Kolibri uses a mixture-of-experts design: about 3.5 billion of its 78 billion parameters are active for each token. But the full model must remain in memory, so its comparatively low per-token computation does not remove the need for substantial hardware.
Aleph Alpha says Kolibri supports context lengths of up to 1,048,576 tokens. The source account says training reached 262,144 tokens and that the longer capability was validated by the company.
The company describes Kolibri as sovereign, citing its development and training under German and European jurisdiction and customers’ ability to run the open weights on their own infrastructure. That framing has limits: the model card says some English training text was rephrased using Google’s Gemma 4, German text using Mistral-NeMo, and Qwen3-32B was used to label data for quality filters.
Benchmark results cited in the source account are Aleph Alpha’s evaluations, not independent verification. The account’s author says they had not run the model themselves. The model card estimates a need for about 78 GB of GPU memory, with all parameters held in memory despite only a fraction being active per token.
Kolibri is positioned for organisations that want to process German or English documents on their own infrastructure. Its hardware requirements and the company-reported results are relevant considerations for institutions weighing control over deployment against operating cost and performance.
