13 October 2026

 

 

For an AI assistant helping to review a German building application, fluent language is only part of the job. It must work with the rules relevant to that project, distinguish the jurisdictions involved and support its explanation with the right documents. For the authority using it, there is another requirement: keeping sensitive information under its own control.

This is the type of issue Heidelberg-based Aleph Alpha is aiming to address with Kolibri, a German-English large language model (LLM) released on 3 October 2026 and announced in a press release on 5 October. Designed for government and industry, it is built to operate on infrastructure customers control. Its language specialisation sits within a broader ambition: giving organisations sovereignty over the AI they use. 

Sovereignty is central to Aleph Alpha's proposition: for them, sovereign AI means retaining control over where a model runs, how information is processed and how the system is adapted. That control becomes tangible as soon as confidential information enters the workflow.

 

Sovereign AI and on-premise deployment

 

For organisations handling confidential material, Kolibri’s deployment model is central to its appeal. Aleph Alpha makes Kolibri’s weights available for download under the Apache 2.0 licence for on-premise use, in other words without sending internal data to external inference services.

That allows an organisation to host the model entirely on its own infrastructure. In a fully internal configuration, prompts, documents and model responses can remain within its security boundary, rather than being submitted to an outside AI provider.

For a manufacturer, this could make it possible to work with proprietary industrial blueprints without transferring them to a third-party model service. For a government body, it creates a route to processing sensitive information internally. 

Organisations can also choose to deploy a downloadable model in a European cloud environment that meets their sovereignty and security requirements. This puts data with an external provider rather than physically inside the organisation. European hosting alone does not establish sovereignty: ownership, legal jurisdiction, access controls and contractual arrangements also need scrutiny.

In either setting, keeping information within the intended boundary requires the surrounding application to be configured accordingly. An internally hosted model connected to an external search service or tool could still send information out. Deployment freedom gives an organisation the ability to control those choices.

That control also connects sovereignty to the bias debate. Downloadable model weights (the learned numerical settings behind its responses) allow organisations to adapt the model for particular workflows. They can choose the documents it draws on, set access rules and evaluate how it treats different requests. These choices create opportunities to correct weaknesses, alongside responsibility for checking the effects of any changes.

Intersting in that regard is that in September 2026, Aleph Alpha has also signed an agreement to combine with Canada’s Cohere. Operating globally as Cohere, the unified company has stated that it will "effectively create the first transatlantic sovereign AI solution." The company will remain deeply rooted in both countries with headquarters, R&D centers, and leadership roles in Canada and Germany.

 

How Kolibri is built for German-language work

 

Control over deployment addresses where information goes. Kolibri’s German-language design addresses another part of the task: how that information is processed and understood. Like other language models, it breaks words and sentences into tokens, the units it uses to learn and generate responses. A tokenizer suited to German can represent its compound words more efficiently, reducing the amount of processing needed for a given passage. The company’s documentation places German at more than a fifth of the pre-training mix, alongside English and code. 

Its technical account also describes a revealing problem earlier in the development process. Data filters configured around English could reject German administrative documents because their average word length was too high. The team adjusted the filters to retain material characteristic of the language.

That is an unusually concrete illustration of how an engineering choice can affect what a model learns. A filter intended to improve data quality could remove exactly the kind of writing a public-sector assistant needs to understand. Assumptions can enter an AI system long before anyone asks it a politically sensitive question.

There is a broader editorial question here, familiar to anyone who works across languages. Translating a text changes its wording; it does not automatically change the institutions, places or experiences it describes. A German translation of an American workplace guide still begins with an American workplace.

Europe itself illustrates the scale of that challenge. A shared continent contains many languages, legal traditions and administrative cultures, alongside substantial differences within individual countries. Familiarity with French institutions offers no guarantee of understanding their Polish or Dutch counterparts.

Representing this diversity requires relevant sources, attention to linguistic subtleties and evaluation in the settings where people will use a system. Kolibri’s German-English focus offers a narrower approach to that task. Its local relevance still needs to be established across different communities and workflows.

 

Language fluency does not settle the AI bias question

 

This need for evaluation is supported by independent research, which separates language capability from cultural representation.

In a 2025 study, Jonathan Hvithamar Rystrøm, Hannah Rose Kirk and Scott Hale compared model responses with World Values Survey opinion data across Danish, Dutch, English and Portuguese. Across the model families studied, they found no consistent relationship between stronger language capability and closer cultural alignment. 

A separate 2024 study led by Tarek Naous examined 16 language models in Arabic tasks, finding preferences for Western cultural entities and examples of stereotyping and cultural unfairness. The researchers also analysed Arabic training corpora and questioned how well commonly used sources represent the relevant cultural contexts. 

Neither study tested Kolibri. Together, however, they explain why fluent German alone would be insufficient evidence that a model represents German-speaking users well.

For an institution choosing an assistant, three questions follow. Can it handle the language? Does it understand the relevant local context? Does it treat different people fairly? A system could perform well on administrative terminology while still making unreliable assumptions about a person’s background or circumstances. Aleph Alpha acknowledges that risk in Kolibri’s model card, warning that it may reproduce systematic and political biases present in its training data. 

 

Document-based answers for government work

 

Testing those risks requires moving beyond a model’s fluency to the work an organisation wants it to support. An official reviewing a building application may need it to distinguish a general requirement from an exception in the supplied guidance. If the necessary local document is missing, the useful response is to identify that gap, rather than supply a plausible rule from elsewhere.

Kolibri supports retrieval-augmented generation, or RAG, in which relevant documents are retrieved and supplied to support an answer. Aleph Alpha says it is trained to abstain when that material contains insufficient evidence. For the authority, this is intended to make the boundary between documented information and an unanswered question more visible; training for abstention does not guarantee that every unsupported answer will be caught.

The same task also exposes the limits of language specialisation. A resident might describe their circumstances in imperfect German or regional expressions, while the official documents use formal administrative prose. An assistant needs to connect those accounts accurately. If it simplifies a letter, it must preserve its conditions and deadlines without making unsupported assumptions about the reader.

These are practical evaluation tasks for a public service: does the system retrieve the right document, retain an important exception and respond appropriately when the evidence is incomplete? The model card places Kolibri on the advisory side of such work: people should review its outputs before acting, and it is not intended to make decisions autonomously. 

 

Sensitive industrial documents and practical benefits

 

Industry offers a related test. Imagine an engineer checking whether an internal specification permits a proposed component change. The assistant needs to find the applicable version, preserve qualifications and make its evidence available for review. Keeping that process internal allows the organisation to use proprietary information without submitting it to an outside model service.

Aleph Alpha’s website describes applications that illustrate the commercial rationale. It reports a 90% reduction in search time through an AI agent working with sensitive documentation at a global chip manufacturer, and 40% time savings in processing requests for quotation at an automotive technology supplier. 

Those examples illustrate the business proposition behind sovereignty: making valuable internal knowledge easier to use while retaining control over its processing. The benefit depends on more than the model. Document access, version management and the way answers are checked all influence whether faster retrieval translates into better decisions.

 

The computing costs of hosting Kolibri

 

Local deployment must also be economically workable. Kolibri uses a mixture-of-experts architecture, activating selected parts of its network for each token. It has roughly 78 billion parameters in total, with approximately 3.46 billion active per token, according to its model card. 

Activating a fraction of the network limits computation per token, but does not remove the need to hold the full model in memory. Aleph Alpha lists a weight-memory footprint of approximately 78 GB in its published format, before the additional memory required to serve requests. Hardware costs, response speed and the number of simultaneous users therefore sit alongside answer quality in the purchasing decision. A system that performs well in a demonstration must also work reliably through a busy day, with enough capacity for the people relying on it.

That brings the story back to the building application. A useful assistant would help the official locate the applicable guidance, understand the applicant’s circumstances and recognise when the evidence is insufficient, while keeping sensitive material within the intended boundary.

Sovereignty gives an organisation more control over how that system is built and operated. Language specialisation aims to make it better suited to the work. Their combined value will be measured in whether people can reach better-supported decisions, using a system whose capabilities and limits they can examine.

 

 

 

Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.

👉 https://buymeacoffee.com/movetheneedle.news