Aleph Alpha Releases Kolibri for Self-Hosted German and English AI

Government agencies and industrial companies have a new downloadable option for German and English AI work. Aleph Alpha released Kolibri 1 on October 3, 2026, with weights available on Hugging Face under Apache 2.0. It is designed to run on customer-controlled infrastructure, without sending internal documents to an outside inference service.

The release targets public administration, industry and aerospace rather than treating every language and workload equally. Aleph Alpha says it built the model in Germany and trained it in Germany and Finland. Its sovereignty pitch combines control over development with customers’ freedom to choose where the model runs.

Open weights do not mean every part of development is open. The Apache 2.0 license covers weights and configuration files; Aleph Alpha retains rights to its training code and methods.

Less computation, not less model to store

Kolibri contains 78.1 billion parameters—the learned values that shape its responses—but activates only 3.46 billion for each token, or chunk of text. Its mixture-of-experts design routes each token through six of 384 specialist components, plus a shared expert. That limits the portion doing work at any moment.

The storage requirement is different. All the weights must remain in memory, even when most are inactive. AI engineer Tejas Kumar’s technical walkthrough puts the weights alone at about 78 GB in eight-bit floating-point format. Running the model also requires Aleph Alpha’s plugin for vLLM, the software used to serve responses.

78.1 billionTotal parameters

The full model contains 78.1 billion parameters.

3.46 billionActive per token

Only 3.46 billion parameters are active for each token.

German depth, with uneven competitive results

German accounts for 21.3% of Kolibri’s pretraining tokens. Aleph Alpha reports 20 trillion pretraining tokens overall and says translation supplied only 6% of the data. A bilingual tokenizer—the system that divides text into chunks—also supports its emphasis on German rather than treating the language as an afterthought.

Kumar tested that tokenizer on Germany’s Basic Law. Kolibri represented the German text in 35,190 tokens, versus 41,482 for the tokenizer used by GPT-5. The English translation produced nearly equal counts. This was a tokenizer experiment, not a model-quality test: fewer chunks let more of the same German text fit into a fixed context window.

In Aleph Alpha’s launch benchmarks, Kolibri scored 87.5 on the German version of the AIME 2025 math test, versus 82.9 for Qwen3.6-35B-A3B. But Qwen led on the overall BFCL v4 tool-calling test, 67.2 to 61.4, and LongBench Pro, 70.8 to 64.5. These are company-run comparisons, not independent validation.

Length also needs qualification. Kolibri’s longest trained context was 262,144 tokens; Aleph Alpha validated it up to 1,048,576, according to Kumar. The larger figure describes tested reach, not the length used throughout training.

Its attention design reduces the work needed for long inputs. Forty of its 50 layers look only at the preceding 512 tokens. Every fifth layer looks across the full preceding text, combining local processing with periodic access to the wider document.

A document assistant, not a trivia champion

Aleph Alpha trained Kolibri to decline answers when supplied documents lack the evidence. Its Merlin-Arthur method alternates examples where supporting evidence remains visible with examples where that evidence is hidden. The model learns either to answer from the document or acknowledge that it cannot.

Users can also choose none, low, medium or high reasoning effort for each request. Aleph Alpha presents those settings as a way to trade response time and cost against answer quality on the same model.

• More willingness to abstain: Aleph Alpha reports Kolibri withheld an incorrect answer on 44% of relevant AA-Omniscience items, compared with 15% for its earlier Kolibri Origin model.
• Less knowledge from memory: Kumar’s account of the company’s evaluation lists 14.8% correct answers on Omniscience, versus 22.2% for Qwen3.5-35B-A3B.