German AI startup Aleph Alpha has unveiled a new language model, Kolibri, targeting heavily regulated fields such as public administration, industry and aerospace.

On Oct. 6, Japanese media outlet Gigazine reported that Aleph Alpha announced Kolibri on Oct. 3, Germany’s Day of German Unity. Kolibri uses a mixture-of-experts architecture and has a total of 78.1 billion parameters, but activates only about 3.46 billion when processing tokens. It supports English and German, and its maximum context length is 1 million tokens.

The company presented security- and compliance-sensitive work such as public administration, manufacturing and aerospace as key markets. It claimed, "In our own benchmarks for math and coding, long-form processing and agent-type tasks, it delivered similar performance to some open-weight models with up to four times as many active parameters." The evaluation, however, is based on Aleph Alpha’s own measurements.

The core strategy is “sovereign AI.” The model was developed in Germany and trained on infrastructure in Germany and Finland, with its in-house pipeline handling everything from data curation to pre- and post-training and optimization. Another feature is that customers can deploy it in their own environments and directly control cost, latency and the level of inference.

It also applied its own Merlin-Arthur training method to curb hallucinations. The method trains the model not to force an answer when it lacks grounds to support a response based on the material provided. The company said, "This increased the grounding of answers in tightly regulated industries."

It also strengthened German-language support. Of 20 trillion total pre-training tokens, 21.3 percent were in German, and it developed a tokenizer dedicated to German and English. It also reduced reliance on translation data.

Kolibri’s weights and configuration files were released on Hugging Face under an Apache 2.0 license. The full training code, architecture and training methods, however, have not been released as open source.