Aleph Alpha Releases 78B Kolibri Open-Weight Model
Aleph Alpha releases a German-English MoE model with 78.1 billion parameters, 3.46 billion active per token and Apache 2.0 licensing.
Aleph Alpha has released Kolibri, a German-English mixture-of-experts model that combines 78.1 billion total parameters with only about 3.46 billion active for each token. The company has published the full weights on Hugging Face under the Apache 2.0 license, making the model available for commercial and self-hosted deployment.
Built for controlled deployment
Kolibri supports adjustable reasoning modes, tool calls and context windows of up to one million tokens, although Aleph Alpha recommends 262,144 tokens for practical operation. The model was trained on roughly 20 trillion tokens and uses a 384-expert architecture, with six experts activated per token. Its knowledge cutoff for both English and German is June 18, 2026.
Aleph Alpha is positioning the release around sovereign AI: organizations can run the model on their own infrastructure, keep sensitive data away from external inference providers and inspect the model’s licensing and supply-chain terms. The company lists minimum hardware requirements ranging from two A100 80GB GPUs to one H200, B200 or B300, depending on the deployment configuration.
Why it matters
Kolibri does not by itself establish that Europe has produced a frontier model matching the strongest global systems. Its bilingual scope is narrower than that of leading general-purpose models, and the full FP8 weights still require substantial hardware. But the release gives European governments and regulated companies a commercially permissive alternative that is explicitly designed for local control. The important signal is less the headline parameter count than the combination of open weights, Apache licensing and a small active-parameter footprint. Those choices could make Kolibri useful as infrastructure for European enterprise and public-sector deployments, even if independent evaluations have yet to confirm its broader competitiveness.
Uncle Cat take: Kolibri’s 3.46B active parameters improve serving economics, but its roughly 78GB weight file still keeps “sovereign” deployment out of reach for ordinary teams.