Kolibri

Kolibri is an English-German language model released on 3 October 2026 by Aleph Alpha, the German AI company, which presents it as a sovereign open-weight model for regulated work such as public administration, industry and aerospace. It is a mixture-of-experts transformer with 78 billion parameters in total, of which about 3 billion are active for each token, and Aleph Alpha says it supports contexts of up to 1 million tokens. The full weights are published on Hugging Face under the Apache 2.0 licence, in an FP8 version with a BF16 base, and the company describes self-hosting with vLLM; no hosted API price is stated. Aleph Alpha says it specialised the model for German, reasoning, mathematics and agentic use, and that it validated its training pipeline first on a smaller model, Kolibri Origin (30 billion total, 3 billion active, 65,000-token context). The announcement compares Kolibri with other post-trained models on a range of English and German benchmarks and places it on the frontier of quality against serving cost; those figures are the company’s own evaluations, so no score is stored here. The point of the design is cost: with only about 3 billion parameters active, the model needs far less compute per token than a dense model of similar total size, which matters for organisations that must run models on their own infrastructure.

Not yet independently verified: Every detail here comes from Aleph Alpha’s own announcement, technical report and Hugging Face repository. No independent press report had been found at 2026-10-04, and all benchmark results quoted by the vendor are its own evaluations. We will update this when it can be confirmed, and remove this note.