Aleph Alpha Releases Kolibri, a 78B-Parameter Sovereign Open-Weight Model

AI Tech Team
•
October 4, 2026
•
👁️ 7 views
🖼️ Featured Image / Generated Result
Aleph Alpha Releases Kolibri, a 78B-Parameter Sovereign Open-Weight Model

Aleph Alpha released Kolibri on October 3, 2026 as a new open-weight English-German mixture-of-experts model focused on sovereign and mission-critical AI workloads. The model has 78.1 billion total parameters but activates only about 3.46 billion parameters per token, combining a large model footprint with sparse computation.

What changed

Kolibri is available with full weights on Hugging Face under Apache 2.0 licensing. Aleph Alpha says the model supports context lengths of up to 1 million tokens and provides four reasoning settings: none, low, medium, and high.

Why the architecture matters

The mixture-of-experts design does not activate every parameter for every token. That allows the model to maintain a large capacity while limiting the amount of computation used for an individual token. For organizations running private infrastructure, sparse models can create an interesting balance between capability and serving cost, although actual efficiency depends on hardware and runtime implementation.

Language and sovereignty

Kolibri is explicitly built around English and German, which makes it particularly relevant to European organizations with multilingual enterprise workloads and data-sovereignty requirements. Open weights also give teams more control over deployment, evaluation, and adaptation than a hosted-only model.

Long-context workloads

The 1-million-token context target is useful for large document collections, codebases, contracts, research archives, and other workflows where repeatedly summarizing or splitting information can introduce information loss. Developers should still test retrieval quality and long-context reasoning on their own data rather than assuming a larger context automatically produces better answers.

Practical takeaway

Kolibri is worth evaluating for teams that need an open-weight model, German-language capability, long context, and the ability to keep inference inside controlled infrastructure. Benchmark memory use, latency, reasoning quality, and licensing fit before moving it into production.

Source: Aleph Alpha — Kolibri Has Landed

← Back to all articles