Kolibri is a new open-weight model released by German AI company Aleph Alpha on October 3, 2026. The launch fell on the Day of German Unity, and in its official blog post "Kolibri Has Landed," Aleph Alpha announced that the full weights are available on Hugging Face under the repository Aleph-Alpha/Kolibri-1, under the Apache 2.0 license, so anyone can download, self-host and use the model commercially. It is not aimed at chatbot leaderboards, but at regulated settings such as public administration, industry and aerospace, where data cannot leave the internal network.
The specs: a large body with a small appetite
| Item | Kolibri 1 |
|---|---|
| Total parameters | 78.1B |
| Active parameters per token | About 3.46B |
| Architecture | MoE, 50 layers, 384 experts |
| Context | Up to 1,048,576 tokens; the vendor recommends staying at or below 262,144 in everyday use |
| Weight size | About 78GB (FP8) |
| License | Apache 2.0 |
The point of a mixture-of-experts design is that total parameters set the knowledge capacity, while active parameters set the compute cost of each step. Kolibri activates only about 3.5B parameters per token, so its serving cost resembles a small model's, yet all 78.1B parameters must sit in memory. The official minimum hardware is two A100 80GB cards, two H100s, or a single H200, B200 or B300: it can run inside a company's own data center, but it is not a laptop model.
German was trained in, not translated in
Aleph Alpha stresses its bilingual design: German makes up 21.3% of the pre-training corpus, machine translation only about 6%, and the team built a tokenizer tailored to German word formation, arguing that translated text carries the cultural fingerprint of its source language. The model was trained on infrastructure in Germany and Finland, on 20 trillion tokens filtered down from more than 200 trillion tokens of raw data.
In the vendor's own evaluations, Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond and 85.9 on LiveCodeBench v6, with a German average of 70.8 that leads comparable open MoE models. Two caveats matter: every number comes from Aleph Alpha itself, with independent replication still pending; and in the same table, the dense Qwen3.8 27B scores higher in German, at roughly eight times the compute per token.
What it is really selling is sovereignty
Aleph Alpha defines sovereignty in two layers. The first is how the model was built: data curation, training and evaluation all took place under European law, designed from the start around the EU AI Act, the General-Purpose AI Code of Practice and the GDPR. The second is how the model reaches customers: open weights mean a customer can move the model into its own facility, and a vendor shutdown cannot affect a system already deployed.
Two details show it is built for production. One is abstention training: with the Merlin-Arthur protocol, Kolibri is trained to say "I don't know" when the context provides no basis for an answer, instead of inventing one, which matters more in government and compliance work than two extra benchmark points. The other is that four reasoning levels (none, low, medium, high) and tool calling are already available, with an official vLLM-based inference plugin, so the deployment path is ready-made.
Three bills to check before you adopt it
The first is memory: starting at 78GB of weights, a pair of 80GB cards is the entry ticket, so check the hardware budget first. The second is language: Kolibri goes deep on German and English only, and Chinese tasks are not its home turf. The third is verification: vendor scores are a reference, not proof, so run your own document set through abstention accuracy and long-document QA before putting it into a regulated workflow. For European institutions, and for teams serving European customers, Kolibri's value is not its parameter count. It is that it bundles three things into one package: data that never leaves the jurisdiction, a model that can be audited, and deployment that no vendor can switch off.
