Aleph Alpha has launched Kolibri, a 78 billion parameter model trained primarily on German and English data to support European digital independence. The system activates approximately three billion parameters per token using a mixture-of-experts architecture and claims to outperform comparable models from March and April 2026 while maintaining lower operating costs. German text constitutes 21.3 percent of the training set, supported by a dedicated pipeline, though Chinese models generated synthetic data as well. Development occurred under the EU AI Act, with training executed on 768 B200 GPUs across Germany and Finland. The weights are distributed via Hugging Face under an Apache 2.0 license.
The release targets public administration, aviation, and industry sectors where data privacy and regulatory compliance are critical. By hosting a large-scale model within Germany and Finland, the project demonstrates that high-performance inference does not require reliance on American cloud infrastructure. This approach provides a functional alternative for organisations seeking to keep sensitive datasets within national borders while adhering to local legal frameworks.
- Training executed on 768 B200 GPUs in Germany and Finland
- Context window capacity reaches one million tokens
- Weights available under Apache 2.0 license on Hugging Face
