Aleph Alpha’s Sovereign A.I. Model Kolibri Is No Match for the Open-Weight Leaders

Set Trending Topics as a preferred source on Google.

Small, fast and German: With Kolibri, Aleph Alpha wants to prove that Europe can still build its own A.I. models. The Heidelberg-based A.I. company has just released the model, timed to Germany’s Unity Day holiday, with full weights on Hugging Face under the permissive Apache 2.0 license. But it is being measured against rivals that, in the fast-moving world of open-weight models, are already six months old.

What Kolibri Can Do

Kolibri is a mixture-of-experts model with 78 billion parameters, of which only about 3 billion are active per token. Its context window reaches up to one million tokens. According to Aleph Alpha, it is designed for public administration, industry and aerospace, meaning customers who do not want to send their data to cloud services in the United States or China and who run the model in their own data centers.

The key facts:

• Training: 20 trillion tokens of pre-training on 768 Nvidia B200 GPUs, nearly 24 trillion tokens in total. Training ran on infrastructure in Germany and Finland.
• German: 21.3 percent of the training data is German, with very little machine translation. Its own tokenizer breaks down German text more efficiently than those of GPT-5, Gemini or Qwen. The word “Bundessozialgerichtes” (Federal Social Court, genitive), for example, becomes “Bundes|sozial|gericht|es.”
• Reasoning: The model can “think” at four levels, from not at all to intensively, to trade cost against quality.
• Hallucinations: Kolibri was trained to say “I don’t know” when the answer is not in the documents.

The pace is remarkable. Its predecessor, Kolibri Origin, with 30 billion parameters, was never released, and only three months passed between the end of its training and that of Kolibri.

Who Kolibri Is Up Against

In its own benchmarks, Aleph Alpha compares Kolibri with three open-weight models: Alibaba’s Qwen3.6-35B-A3B, Nvidia’s Nemotron 3 Super 120B-A12B and Mistral Small 4 119B-A6B from France. In that field, Kolibri holds up well. It leads the AIME 2025 and 2026 math tests with 96.9 and 96.0 percent, the GPQA Diamond knowledge test with 84.3 percent and the LiveCodeBench coding benchmark with 85.9 percent. On agentic tasks such as τ²-bench Telecom and the BFCL tool-calling test, Qwen3.6 comes out ahead, as it does on long contexts. Aleph Alpha says Kolibri keeps up with rivals that have up to four times as many active parameters.

The catch: All three comparison models date from the spring. Since then, Chinese labs in particular have released considerably stronger open-weight models, including the Qwen3.8 family, Z.ai’s GLM-5.3, Moonshot AI’s Kimi K3 and Xiaomi’s MiMo-V2.6-Pro. Kolibri is not compared with any of them.

Where Kolibri Would Land on Artificial Analysis

There is no independent ranking yet; Kolibri is not yet listed on Artificial Analysis. An estimate is still possible through the comparison models. In the current Artificial Analysis Intelligence Index, Qwen3.6-35B-A3B scores 18 points, Nemotron 3 Super 13 and Mistral Small 4 11. If Kolibri performs roughly at the level of Qwen3.6, as Aleph Alpha’s benchmarks suggest, a score of around 15 to 20 points would be realistic.

That would put Kolibri behind at least 20 other open-weight models. Here is what the top of the field currently looks like:

The giant frontier models are hardly a fair opponent for a model with 3 billion active parameters. Yet even in its own weight class there is now stronger competition: K2 Horizon MoVA, with 4 billion active parameters, scores 25 points, and Alibaba’s Qwen3.8-Flash-Next reaches 40 with 6 billion active parameters.

Why Aleph Alpha Sees It Differently

Aleph Alpha argues that public benchmarks do a poor job of reflecting what its customers need. The company has therefore built its own test suites for sectors such as public administration, automotive suppliers and aerospace. There, Kolibri has improved markedly over its predecessor, in aerospace from 0.14 to 0.59 points. The company also says Kolibri leads the compared models on quality relative to serving cost in both German and English. And its central selling point is sovereignty anyway: a model developed in Germany under European law, built with the E.U. A.I. Act and the G.D.P.R. in mind, that customers can run entirely on their own.

How long the Aleph Alpha brand will pursue that path is an open question. The company recently signed a binding agreement to merge with the Canadian A.I. company Cohere. The combined company is to operate under the Cohere name, with Heidelberg remaining a research hub.