Aleph Alpha launches Kolibri, a German-trained AI model, weeks before Cohere deal closes

Aleph Alpha has released Kolibri, an open-weight language model built and trained entirely in Europe, as the Heidelberg company tries to prove German AI can still compete on the world stage. The model went live on 3 October and was formally announced two days later, arriving just weeks after Aleph Alpha agreed to merge with Canada's Cohere.

The model uses a mixture-of-experts design, carrying 78 billion parameters in total while switching on only about 3.5 billion of them for any single query. That design lets it run faster and cheaper than a dense model of similar size, which Aleph Alpha says matters for customers who want to run the model on their own servers rather than send data to a cloud provider in the United States or China.

Built around German, not bolted on

Roughly a fifth of Kolibri's training data is German-language text, and the company built a custom tokenizer that breaks German words into their component parts rather than treating long compound nouns as single, inefficient chunks. Aleph Alpha trained the model to reason in German natively rather than translating internally from English.

The model supports retrieval-augmented generation, tool calling and multi-step agentic tasks, and it has an adjustable reasoning setting that lets users trade response speed for depth. Aleph Alpha says it specifically trained Kolibri to say it does not know an answer when the documents it is given do not contain enough evidence, a feature aimed at government and industry users who cannot tolerate confidently wrong answers.

"Kolibri demonstrates that we have the talent and expertise in Germany to develop competitive AI models," said Ilhan Scheer, chief executive of Aleph Alpha. "For us, AI sovereignty means freedom of choice by retaining the ability to build and advance this technology, and giving customers control over how they use it."

Samuel Weinbach, the company's co-founder and co-chief research officer, framed the model as a response to a specific gap. "We designed Kolibri around the needs of German-language applications, from a tokenizer optimized for German to training the model to reason in German," he said.

Compliance as a selling point

Aleph Alpha is leaning on paperwork as much as performance. The company published a technical report detailing how it screened training data against a blocklist of more than 4.5 million URLs, including sources tied to the European Commission's piracy watch list, and says every third-party dataset was checked for licensing terms and opt-out compliance before use. That level of documentation is meant to help customers in regulated industries show they meet EU AI Act and data-protection requirements, something that is harder to prove with closed, US-trained models.

The model's full weights are published on Hugging Face under the permissive Apache 2.0 licence, meaning businesses and public bodies can inspect, modify and self-host it without paying for access. Aleph Alpha positions that openness, combined with training entirely on European infrastructure, as the core of its sovereignty pitch to government agencies and companies in sectors such as aerospace and automotive supply.

A narrow release window

Kolibri lands at an awkward moment for the brand that built it. In September, Aleph Alpha signed a binding agreement to combine with Cohere, forming what the two companies describe as a transatlantic sovereign AI group that would operate under the Cohere name, with Heidelberg kept on as a research site. The deal still needs regulatory sign-off, and until it closes Aleph Alpha continues to operate as an independent company.

That timing raises an obvious question for European buyers evaluating the model: how long will an Aleph Alpha-branded product line exist independently of Cohere's roadmap. The company has not said how Kolibri fits into the combined entity's future plans.

On raw capability, outside reviewers have pointed out that Aleph Alpha's own benchmark comparisons use rival open-weight models released earlier in the year, rather than newer competitors from Chinese labs that have since posted stronger results on public leaderboards. Aleph Alpha counters that its sector-specific test suites, covering public administration, automotive and aerospace use cases, show larger gains over its unreleased predecessor model than generic benchmarks capture.

For European buyers who have long been told to wait for a credible homegrown alternative to US and Chinese model providers, Kolibri is a concrete answer, even if an incomplete one. Whether it keeps its name on the box once the Cohere merger closes is a separate question that Aleph Alpha has yet to address publicly.