Aleph Alpha open-sources Kolibri, a German-English MoE for sovereign AI

Aleph Alpha just shipped Kolibri, an open-weight German and English language model aimed at sovereign, on-premise use in government and industry. The Heidelberg company timed the release for 3 October 2026, Germany's Day of German Reunification, and put the full weights on Hugging Face under Apache 2.0.

What Kolibri is

Kolibri is a Mixture-of-Experts Transformer with 78.1 billion total parameters and about 3.46 billion active per token. That split is the point: enough capacity for serious work, small enough to serve efficiently on hardware you control. Aleph Alpha says the model supports context lengths up to 1 million tokens (trained natively to 262,144, with a path to extend further). It ships with four reasoning effort levels (none, low, medium, high) and tool calling.

The company positions Kolibri for mission-critical work in regulated areas such as public administration, industrials, and aerospace. Customers are meant to run it in their own infrastructure rather than send sensitive data to a third-party chat API.

How it was built

Aleph Alpha trained Kolibri on infrastructure in Germany and Finland under European and German law. German made up about 21.3% of the pre-training mix. The team first validated a smaller sibling, Kolibri Origin (about 30B total, 3B active, 65k context), then scaled the same pipeline to Kolibri in roughly three months between June and September 2026 pre-training finishes.

The public tech write-up leans hard on sovereignty: transparent weights, documented data curation, and freedom to deploy without handing IP or compliance to a foreign cloud. Benchmarks in the blog put Kolibri on a strong quality-versus-serving-cost curve for both English and German against other post-trained MoE models in a similar active-parameter band.

Why it matters

Europe keeps asking for "sovereign AI" that agencies and factories can run locally. Kolibri is a concrete open-weight answer from a German lab, not another closed US frontier API. Open weights plus Apache 2.0 mean enterprises can inspect, fine-tune, and host the model themselves. The bilingual design matters too: German admin and industrial prose are not just translated English, and Aleph Alpha says it optimized the tokenizer and data mix for that.

Dany's take

I care about this because on-premise German and English models are still scarce at this quality band. A 78B MoE with only ~3.5B active is the kind of shape that actually fits real servers. Whether Kolibri becomes a default for German public sector and industry depends on how well the evals hold up outside Aleph Alpha's own harnesses, and on support after Cohere's announced deal with the company. For now, an Apache 2.0 release on Hugging Face is the right kind of move.

Source: Aleph Alpha: Kolibri Has Landed.