Aleph Alpha's own post, dated 3 October 2026, announces Kolibri, an English-German mixture-of-experts model released with open weights. The numbers below are the vendor's own, run on its own evaluation set-up, and not independently verified.
As stated by the vendor
- 78 billion total parameters, with about 3 billion active per token.
- Apache 2.0 licence; weights published on Hugging Face and a serving package with a vLLM plugin.
- Pre-trained on 20 trillion tokens, with German making up 21.3% of the pre-training tokens.
- Context length of up to 1 million tokens is claimed; the specification table lists 262,144 as the longest trained length.
- Vendor benchmarks include 96.9 on AIME 2025 and 84.3 on GPQA Diamond.
Why it matters
A permissively licensed model with a small active-parameter count can run on modest GPU hardware, which makes it a candidate for on-premise or sovereign deployments where data cannot leave the organization. Whether it suits you depends on your tasks.
Do not choose a model from vendor benchmarks. Test it on your own examples, in your language, against the alternatives.