Aleph Alpha releases Kolibri, an open German-English AI model for regulated sectors

Aleph Alpha has released Kolibri, an open-weight language model designed for German and English work in regulated organizations. The company positions the model for public administration, industry, aerospace, banks, and other users that want to operate AI systems on their own infrastructure.

Kolibri uses a mixture-of-experts architecture. It contains 78.1 billion parameters in total, but activates about 3.46 billion for each token, the pieces of text that AI models process. This approach aims to reduce computing costs during use while retaining the capabilities of a much larger model. The full model still needs to remain in GPU memory, however. Aleph Alpha says deployments require roughly 78 GB of GPU memory.

The model weights and configuration files are available under the Apache 2.0 license. This allows organizations to download, modify, fine-tune, and run the model themselves. Aleph Alpha retains rights to its training code and methods.

In its official announcement, Aleph Alpha describes Kolibri as a “sovereign” model. The term refers both to its development and deployment. According to the company, teams in Germany built the model, and it was trained on infrastructure in Germany and Finland under European and German law. Customers can also run the weights in their own environment, rather than sending internal documents to an external inference service.

Built for German documents and long contexts

Kolibri supports a native context window of 262,144 tokens and has been tested at up to 1 million tokens. That could make it useful for tasks involving long contracts, technical manuals, legal files, or collections of internal documents.

Aleph Alpha also developed a bilingual tokenizer, the component that breaks text into processable units. The company says its UniBPE tokenizer handles German compound words more efficiently than several competing tokenizers. In an independent experiment described by Tejas Kumar in his analysis, Kolibri required substantially fewer tokens than OpenAI’s o200k tokenizer for the German Basic Law, while producing a similar token count for the English translation.

The model was trained on almost 24 trillion tokens across several training stages. German represented 21.3 percent of its pretraining data, according to Aleph Alpha. The company says this was intended to avoid treating German as a secondary language added to an English-first model.

Kolibri also offers four reasoning settings: none, low, medium, and high. Users can select how much processing effort the model should apply, trading response speed and cost against more extensive reasoning.

Grounding is a central claim

Aleph Alpha emphasizes that Kolibri should abstain when supplied documents do not support an answer. The company trained this behavior through its Merlin-Arthur protocol, which creates examples where relevant evidence is either present or deliberately removed. The goal is to teach the model to distinguish between supported answers and unsupported guesses.

On the company’s reported tests, Kolibri abstained or gave a partial answer on 44 percent of questions it could not answer correctly in an Omniscience benchmark. That does not mean the model eliminates hallucinations, but it may be valuable in retrieval-augmented generation systems, where a model answers based on supplied company documents.

The published results also show limitations. Kolibri trails some competitors on closed-book knowledge questions, multi-turn tool use, and coding-agent benchmarks. Its long-context performance reportedly improves at the 1 million-token range, but it lags one competitor on a test at 128,000 tokens. The model currently requires Aleph Alpha’s vLLM plugin, adding another operational requirement for organizations that want to deploy it.

For teams handling sensitive German-language content, Kolibri offers an open alternative focused on local deployment, document grounding, and long inputs. Its practical value will depend on independent testing and on how well it performs with each organization’s own data and workflows.

Sources

• Show HN: Germany’s new sovereign AI model Kolibri – Tejas Kumar
• Kolibri: A Sovereign Open-Weight Model – Aleph Alpha