Aleph Alpha's Kolibri Open-Weight Model Ships Free

Kolibri open-weight model went live Oct. 3, with German AI lab Aleph Alpha releasing the 78 billion parameter system on the Day of German Reunification, framing the launch as a statement on European technology independence.

Key Takeaways

Aleph Alpha released the 78 billion parameter Kolibri model on Oct. 3 under the Apache 2.0 license
Kolibri supports context windows up to 1 million tokens, processing roughly 750,000 words in a single pass
Only 3 billion of Kolibri’s 78 billion parameters activate per query through its Mixture-of-Experts architecture
Aleph Alpha has not published independent benchmark comparisons between Kolibri and frontier US models

The model is free to download on Hugging Face under the Apache 2.0 license, meaning any company or government can run, modify or resell it without paying Aleph Alpha a fee. Kolibri supports context windows up to 1 million tokens, letting it process roughly 750,000 words of text in a single pass.

Kolibri is a Mixture-of-Experts transformer, an architecture that routes each query to a small subset of specialized sub-networks rather than activating the full model every time.

Of its 78 billion total parameters, only 3 billion activate per query, a design choice that keeps inference costs low while retaining a large overall knowledge base.

Aleph Alpha said the model is tuned for “sovereign mission-critical work” in public administration, industrial and aerospace settings, sectors where European governments are wary of depending on US-hosted AI infrastructure.

Also Read: AI Regulation Approved To Ban Solo AI Job Decisions In California

The Pipeline Behind The Release

Kolibri is not Aleph Alpha’s first attempt. The company built Kolibri Origin first, a smaller 30 billion parameter model with a 65,000-token context window, to validate its training pipeline before scaling up.

That earlier model let engineers run hundreds of ablation experiments, tests that isolate which training ingredients actually improve performance, and stabilize pre-training runs without manual intervention when hardware failed.

Aleph Alpha has positioned itself for years as Europe’s answer to OpenAI and Anthropic, arguing that governments and regulated industries need AI they can audit and host domestically rather than renting from US hyperscalers.

The company recently signed an agreement with Cohere to build what it called the first transatlantic sovereign AI offering, a partnership aimed at giving European and Canadian institutions a non-US alternative for frontier-grade models.

What Sovereignty Buys, And What It Costs

The open-weight release matters because it tests whether a mid-size European lab can compete on raw capability, not just on data residency promises.

Kolibri’s 1-million-token context window matches or exceeds many commercial rivals, but Aleph Alpha has not published independent benchmark comparisons against frontier US models. Enterprise customers in regulated sectors will decide whether German-trained weights are worth trading off against the scale of OpenAI or Google‘s training budgets.