Aleph Alpha has released Kolibri, a new open‑weight language model positioned as a sovereign AI solution for mission‑critical environments. Kolibri is an English‑German Mixture‑of‑Experts Transformer with 78 B total parameters and 3 B active parameters per token. The model can handle context lengths of up to 1 M tokens and is distributed under the Apache 2.0 license via Hugging Face.
The opening image features the Kolibri hummingbird logo against a green gradient, representing the model's brand identity as a sovereign AI solution.
The release emphasizes sovereignty through two dimensions: how the model was built and how it transfers to customers. Aleph Alpha provides full supply‑chain integrity, documenting every step from data ingestion through pre‑ and post‑training to final evaluation. The company states that customers enjoy freedom of deployment and intellectual‑property safety, inheriting compliance as a property of the model itself.
Kolibri was specialized for German, reasoning, math, and agentic behavior. A bilingual German/English tokenizer was developed, and 21.3 % of the pre‑training tokens are organic German data, meaning the model is bilingual by design rather than an English model that merely read some German. Training was conducted on infrastructure in Germany and Finland under European and German law, with no foreign control.
The menu screen shows the Kolibri model options, reflecting the model's accessibility for regulated sectors and its bilingual German/English tokenizer.
Benchmark results place Kolibri on the Pareto frontier for quality versus serving cost in both English and German. In many tasks—including math, coding, and long‑context reasoning—Kolibri matches models with up to four times its active parameter count, such as Nemotron 3 Super. The model also inherits abstention training (the Merlin‑Arthur protocol), enabling it to output "I don’t know" when the context does not support an answer, a feature valued by customers.
Kolibri Origin, a prior 30 B model, served as an intermediate step; Kolibri improves upon it with more parameters, a longer context window (up to 256 k tokens trained vs 65 k for Origin), and a larger training token count (20 T vs 7.5 T). The training pipeline was implemented as GitHub Actions workflows, automatically handling hardware interruptions and checkpoints.
The panel features Aleph Alpha team members discussing Kolibri's development and sovereign AI positioning at the public release.
Overall, Kolibri aims to give regulated sectors a transparent, deployable model whose size and efficiency allow on‑premise operation without sending internal data to third‑party services.