CONFIRMED: Aleph Alpha launches Kolibri, giving Europe its first open model that truly competes
On the day Germany celebrates its reunification, Aleph Alpha, an AI lab based in Heidelberg, published Kolibri: a German and English bilingual language model built with a Mixture of Experts architecture that totals 78.1 billion parameters while activating only 3.46 billion per processed token. The model supports context windows of up to 1 million tokens, a controllable reasoning mode with different intensity levels, tool calling, and full weights released on Hugging Face under the Apache 2.0 license. This is not a marketing teaser: Aleph Alpha itself published the model card, a detailed technical report, and the files needed to run the model, which makes this launch a confirmed fact, not a rumor.
Why a German company shipping open weights is news
The race for open weight models that actually matter has mostly been run by labs in the United States, including OpenAI's own gpt oss, and in China, with families like Qwen and DeepSeek. Europe, apart from France's Mistral, had largely been watching from the sidelines. Kolibri's technical report places the model on the Pareto frontier of quality versus serving cost, alongside open competitors such as Qwen3.6, Gemma 4 and gpt oss 120b, in both English and German. That does not mean Kolibri beats all of them on every test, but it puts a German lab, for the first time, in a direct and public comparison against the leading names in open models.
What changes in practice for organizations handling sensitive data
For public administrations, regulated industries, and sectors such as aerospace in Europe, the Apache 2.0 license means the ability to download, run, and fine tune Kolibri on an organization's own infrastructure, without depending on an API hosted outside the continent. Because the model activates only 3.46 billion of its 78 billion parameters per token, serving Kolibri tends to cost less than a dense model of the same size, though it still requires serious hardware: Aleph Alpha itself recommends at least two H100 GPUs or equivalent to run the FP8 version. That combination of sovereignty and controlled cost is the launch's central commercial argument, more than any single benchmark score.
The fine print behind the open license
The Apache 2.0 grant covers only the weight and configuration files published in the repository; Aleph Alpha states clearly in the model card that it retains all rights to the training code, architecture, and methods used to build Kolibri. In other words, the model is open to run, fine tune, and redistribute, but the pipeline that created it stays closed, which classifies Kolibri as open weight rather than a fully open source project. Another detail worth noting is context: native training covers up to 256,000 tokens, and the card itself recommends staying within that limit for complex tasks; the 1 million token figure comes from extrapolation, not from the original training run.
MaxAssistant's take
The most interesting detail about Kolibri is not winning or losing any single benchmark, it is the decision to publish, alongside the model, a nearly 200 page technical report detailing data provenance, energy consumption, and the training process. That turns transparency itself into part of the product, directly answering the question that dominates public and corporate procurement in Europe: where does the data go. My take is that Kolibri's real value is not beating Qwen or Gemma on a leaderboard, it is proving that a credible sovereign alternative can exist at all, right as the European Union's AI Act is pushing institutions to ask exactly that question more forcefully. The real test is not the parameter count, it is whether public administrations and regulated companies actually adopt Kolibri at scale.
Sources
Aleph Alpha: official announcement of the Kolibri launch; Aleph Alpha: Kolibri model page and specifications; Aleph Alpha: full Kolibri technical report.