Blog
Aleph Alpha releases Kolibri, a 78B open-weight German-English model under Apache 2.0
On 3 October 2026 Aleph Alpha released Kolibri, a German-English mixture-of-experts model with 78B total and about 3.46B active parameters, a context window of up to 1M tokens, and open weights under Apache 2.0. It is built for organisations that want to run AI on infrastructure they control.
German AI company Aleph Alpha released Kolibri on 3 October 2026, the Day of German Unity. It is an open-weight language model focused on German and English, with full weights on Hugging Face under the Apache 2.0 license.
Aleph Alpha pitches it as "sovereign" AI: a model that governments and companies can run on infrastructure they control, with documented data provenance.
The key specs
From the Hugging Face model card and Aleph Alpha's technical report:
- Architecture: Mixture-of-Experts (MoE) transformer
- Size: 78.1B total parameters, about 3.46B active per token (roughly 4.4%)
- Languages: German and English
- Context: up to 1,048,576 tokens. Aleph Alpha recommends 262,144 or fewer for serving efficiency and complex tasks
- Features: an explicit reasoning mode and tool calling
- Knowledge cutoff: 18 June 2026
- License: Apache 2.0, with FP8 and BF16 weight versions published
Because only a small share of the parameters run for each token, a MoE model like this can cost much less to serve than a dense model of the same total size. It still needs enough memory to hold all 78B parameters.
How it was trained
The technical report says Kolibri was trained on 24 trillion tokens, with German making up more than 20% of the data mix. To reach that share, Aleph Alpha added more than 2 trillion German tokens that it curated from the web or generated synthetically. Post-training used supervised fine-tuning plus reinforcement learning on more than 1.2 million internally curated tasks covering reasoning, tool use, instruction following, code and retrieval.
Aleph Alpha says Kolibri sits "on the Pareto frontier of quality and serving cost" in English and German among the open models in its report. That is the company's own evaluation and has not been independently confirmed.
Running it
Kolibri needs Aleph Alpha's aleph-alpha-inference package, a vLLM plugin, available as a container image or installable package. The model card gives a standard vllm serve command with Kolibri-specific reasoning and tool-call parsers.
The model card also notes that Aleph Alpha has signed the EU General-Purpose AI Code of Practice. That matters now that the EU AI Office can enforce general-purpose model rules.
Business context
In a 5 October press release, Aleph Alpha says the model targets public administration and industry. It also notes that the company recently agreed to a deal with Cohere to build a "transatlantic sovereign AI company." The deal still needs regulatory approval, and Aleph Alpha says it operates independently until then.
Why it matters
Most leading models are American or Chinese and are often available only through APIs. A permissively licensed, long-context European model gives organisations with strict data rules, including those outside Europe, another option they can host themselves.
The trade-offs: it is tuned for German and English, not other languages; it needs a specific inference plugin; and benchmark claims come from the vendor.
What to watch
- Independent benchmarks of Kolibri against other open-weight models
- How hard it is to deploy outside Aleph Alpha's own tooling
- What happens to Kolibri's roadmap if the Cohere deal closes