Kolibri: A Sovereign Open-Weight Model
Kolibri is an English-German Mixture-of-Experts model with 78B total and 3B active parameters, a context of up to 1M tokens and open weights under Apache 2.0, specialized for sovereign, mission-critical work.
Kolibri is a new English-German Mixture-of-Experts model with 78B total parameters and 3B active parameters, designed for sovereign, mission-critical work in regulated sectors. It supports up to 1M tokens of context and is available under the Apache 2.0 license.
What happened
Kolibri is a new English-German Mixture-of-Experts model developed by Aleph Alpha, with 78B total parameters and 3B active parameters. It supports up to 1M tokens of context and is available under the Apache 2.0 license. The model is designed for sovereign, mission-critical work in regulated sectors and is optimized for performance across various domains.
Why it matters
Kolibri's focus on sovereignty, compliance, and performance makes it particularly relevant for enterprises and government agencies that require secure and reliable AI solutions. Its efficient design and open weights also make it accessible for a wide range of applications.
- Kolibri is optimized for performance and cost-efficiency in regulated sectors.
- Its open weights and Apache 2.0 license promote transparency and accessibility.
- The model's focus on sovereignty and compliance makes it suitable for mission-critical applications.
- Kolibri's training pipeline enables rapid development and iteration.
- The model's bilingual design supports both English and German use cases.
The bigger picture
Kolibri's design allows for efficient deployment on-premise, reducing reliance on third-party services. Its focus on compliance and transparency makes it suitable for regulated industries, while its performance across various benchmarks highlights its versatility.
Kolibri represents a significant advancement in sovereign AI development, combining performance, compliance, and transparency to meet the needs of regulated industries and mission-critical applications.
Technical details
- Mixture-of-Experts (MoE) architecture with 3B active parameters out of 78B total.
- Supports up to 1M tokens of context.
- Built with a bilingual German/English tokenizer and 21.3% German pre-training tokens.
- Trained on 768 B200 GPUs with 24T tokens of data.
- Uses exact quantile balancing for expert routing during training.
FAQ
What is Kolibri?
Kolibri is an English-German Mixture-of-Experts model with 78B total parameters and 3B active parameters, designed for sovereign, mission-critical work in regulated sectors.
What are the key features of Kolibri?
Kolibri supports up to 1M tokens of context, is available under the Apache 2.0 license, and is optimized for performance across various sectors with a focus on compliance and transparency.
How was Kolibri developed?
Kolibri was developed using a high-velocity training pipeline that enabled rapid iteration and efficient model training, with a focus on transparency and compliance.
What are the technical specifications of Kolibri?
Kolibri has 78B total parameters, 3B active parameters, and supports up to 1M tokens of context. It uses a bilingual German/English tokenizer and was trained on 768 B200 GPUs with 24T tokens of data.
What are the use cases for Kolibri?
Kolibri is designed for sovereign, mission-critical work in regulated sectors such as public administration, industrials, and aerospace, with a focus on compliance, transparency, and performance.
This summary and analysis were generated with AI from the original article and may contain errors. Read the original for the full story.
Read the full story
Originally published by Lobsters on Oct 4, 2026.
Comments
Sign in to join the discussion.