The launch
On German Unity Day, October 3, 2026, Germany's Aleph Alpha released Kolibri, an English-German open-weight language model under Apache 2.0. It's a Mixture-of-Experts with 78B total parameters and about 3B active per token, and context up to 1 million tokens.
It's pitched as a "sovereign" model for mission-critical work in regulated sectors: public administration, industry and aerospace. Per the company, it was built in Germany, trained on infrastructure in Germany and Finland, under European law with no foreign control, designed from the start around the EU AI Act, the GPAI Code of Practice and GDPR.
The numbers (self-reported)
In Aleph Alpha's own tests at the highest reasoning effort:
- AIME 2025: 96.9 (87.5 in German); GPQA Diamond: 84.3.
- LiveCodeBench v6: 85.9; SWE-Bench Verified: 66.4.
- τ³-bench banking: 38.1, well above same-size rivals.
- The company says it matches models with up to four times more active parameters, such as Nemotron 3 Super.
But it doesn't lead everywhere: Qwen3.6-35B-A3B wins on τ²-bench telecom, BFCL and LongBench Pro, and dense Qwen3.8 27B is ahead on the overall average. Benchmarks were run with the company's own harness.
What's different
- Real German: 21.3% of pre-training tokens are German, nearly all organic or rephrased from German text; translation was kept minimal to avoid "German that sounds American." A bilingual tokenizer compresses German better than GPT-5 and Gemini, lowering inference cost.
- Knowing when to say "I don't know": abstention training with the Merlin-Arthur protocol. On AA-Omniscience Kolibri abstains instead of answering wrong on 44% of items, vs. 15% for its predecessor.
- Four reasoning levels (none, low, medium, high) to trade cost for quality.
- Efficiency: only 10 of 50 layers use full attention; per the company it serves 18 concurrent 256k-token requests on two H100s.
Why it matters
Europe wants alternatives to US and Chinese models for government and industry. An open model small enough to run on-premise, with data and compliance documentation, answers exactly that. The idea travels: Japan is also debating sovereign AI.
For businesses
- Sensitive data on-prem: 3B active means low serving cost on your own GPUs.
- Test on your domain: both public and internal benchmarks are the company's own.
- Grounding matters: in enterprise RAG, a model that abstains beats one that guesses.
- Technical requirement: it needs the aleph-alpha-inference package (a vLLM plugin).
Sources
- Aleph Alpha, "Kolibri Has Landed" (2026-10-03): https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/
- Hugging Face, Aleph-Alpha/Kolibri-1: https://huggingface.co/Aleph-Alpha/Kolibri-1
Want to evaluate open, sovereign models to run on your own infrastructure? KGA IT offers a free consultation: kga-it.com.