Aleph Alpha announced the release on October 3 under the headline "Kolibri Has Landed: A Sovereign Open-Weight Model." Kolibri is German for hummingbird — a fitting name for what the Heidelberg lab describes as a fast, efficient model: a mixture-of-experts (MoE) design that activates only a fraction of its 78 billion parameters for each token it reads or writes. The weights are publicly available on Hugging Face under the Apache License 2.0.

A technical breakdown published the same day puts the numbers in perspective: roughly 78 billion parameters total, with only about 3.5 billion of them engaged per token — under 5% active compute, a ratio that makes Kolibri cheap to run relative to its total capacity. The same write-up notes the model was trained from scratch on infrastructure in Germany and Finland, not fine-tuned from an existing US or Chinese base. The lab positions it for language processing, text generation, reasoning, and agentic tasks.

Visualization of a mixture-of-experts neural network with only a few glowing expert clusters active
Kolibri's mixture-of-experts architecture: 78B parameters total, about 3.5B active per token. AI-generated illustration.

Sovereignty first#

The release lands at a hinge moment for Aleph Alpha itself. The company — founded in 2019 in Heidelberg — signed a definitive combination agreement with Canada's Cohere in September 2026, valuing the combined company at roughly $20 billion and aiming to become, in the companies' words, the first "transatlantic sovereign AI solution." The transaction, still subject to regulatory approval, will dual-headquarter the firm in Berlin and Toronto and keep Heidelberg as a research center. Schwarz Group, Aleph Alpha's anchor investor, committed $600 million to Cohere's upcoming Series E.

That context is the story. Aleph Alpha was an early flag-bearer for European AI — raising a much-hyped €500 million round in 2023 (later reporting showed the genuine equity was closer to €110 million) — then drifted away from foundation models toward enterprise deployments. Kolibri marks a return: a from-scratch, openly licensed European model at a scale where most labs keep the weights closed. Under Apache 2.0, anyone can download, fine-tune, and commercialize it — including the governments and regulated industries the company has always courted.

A European data center hall with a glowing holographic outline of Europe above the server racks
Trained in Germany and Finland: Kolibri's pitch is sovereignty through infrastructure, not just licensing. AI-generated illustration.

Why it matters#

Europe's sovereignty debate has mostly been about deployment — where the servers sit, whose laws apply. Kolibri attacks the other side of the equation: open weights, built locally, at meaningful scale. If the model performs — and that verdict will take days of community benchmarking, which we have not seen yet — it gives European enterprises a serious alternative to fine-tuning American or Chinese bases behind their own firewalls.

The open question is everything else. No verified benchmark figures had surfaced in the sources we read on launch day; treat any scores floating around as unconfirmed until independent evaluations run. And a $20-billion merged company will still need to prove it can outrun US labs on cadence, not just licensing. But as an opening move, this is the most substantive open-weight release out of Europe in a while — and the Apache 2.0 license means the community, not the press release, gets the final say.