Aleph Alpha has released the German-English model Kolibri with publicly available weights
Kolibri from Aleph Alpha has 78 billion parameters, with approximately three billion active per token. The weights are available on Hugging Face under the Apache 2.0 license. According to the technical report, it supports a context of up to one million tokens.
Aleph Alpha has released Kolibri, a model designed for German and English. The model uses a mixture-of-experts architecture: of its total 78 billion parameters, approximately three billion are active when processing a single token. The weights are publicly available on Hugging Face under the Apache 2.0 license.
German accounts for 21.3 percent of the training data, and the company built a dedicated system for preparing German data for the project. Chinese models were also used to generate synthetic training data. According to the technical report, the model supports a context of up to one million tokens and was trained in Germany and Finland on 768 B200 GPUs.
According to Aleph Alpha, the model is focused on public administration, aviation and industry. The company claims that it offers a favorable balance between quality and operating costs and outperforms the models used for comparison with similar architectures on one of these measures. However, according to the article, some of the models used for comparison are significantly older.
Why it matters
Publicly available weights and the Apache 2.0 license allow interested users to assess the model for their own German and English tasks. The claimed context of up to one million tokens is relevant to working with extensive texts; however, the claim about quality and operating costs does not in itself demonstrate a benefit for a specific deployment.
Release card
Kolibri
Aleph Alpha
- Specifications
- The model has 78 billion parameters, of which approximately three billion are active when processing a single token.
- Context
- According to the technical report, the model supports a context of up to one million tokens.
- Licence
- Apache 2.0
- Availability
- The model weights are publicly available on Hugging Face.
- According to the company, the model is intended for public administration.
- According to the company, the model is intended for aviation.
- According to the company, the model is intended for industry.
Aleph Alpha claims that, in German and English, Kolibri outperforms the models used for comparison with similar architectures on either quality or operating costs. According to the article, some of the models used for comparison are significantly older.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
Two audiences, two different impacts
What this means
For individuals
Developers working with German and English gain access to the weights for Kolibri for their own experiments.
For a business
Organizations in public administration, aviation and industry have another model that, according to the company, is designed for their work. They can take the Apache 2.0 license into account when choosing a solution, but the claimed balance between quality and operating costs requires verification on their tasks.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 1 independent. We count feeds from the same owner only once.