Kolibri
Aleph Alpha released Kolibri, an English-German mixture of experts model with 78B total parameters and about 3B active, under Apache 2.0 with weights on Hugging Face.
Kolibri is a mixture of experts transformer for English and German with a context of up to 1M tokens. It was trained from scratch after Aleph Alpha first validated its training pipeline on Kolibri Origin, a 30B total, 3B active model with a 65k context. Aleph Alpha specialized it for German, reasoning, math and agentic work in regulated sectors such as public administration, industrials and aerospace.
- Date
- Saturday 3 October 2026
- Lab
- Aleph Alpha
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Total parameters | 78B 3B active per token per Aleph Alpha; a third-party post says about 3.5B active | company |
| Context length | 1M tokens up to 1M tokens per Aleph Alpha | company |
| German score | 70.8 vs 69.8 for Qwen3.5 35B-A3B; dense Qwen3.8 27B scores 79.9; Aleph Alpha's own evaluation | company |
A trendingtopics.eu headline says Kolibri trails the open-weight leaders, and the third-party post notes a larger dense Qwen model scores higher in German.
Sources
- aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/
- tej.as/blog/aleph-alpha-kolibri
This record was checked and corrected against its sources on 8 October 2026. How we check