AIToday
GIGAZINE AIPublished: Oct 6, 2026, 10:01 JST

Aleph Alpha releases Kolibri, a sovereign open-weight model

Aleph Alpha releases Kolibri, a sovereign open-weight model

3 Key Points

  1. What happened

    On October 3, 2026, German AI startup Aleph Alpha announced Kolibri, an open-weight language model with 78 billion total parameters but only 3 billion active ones per token. It supports English and German up to 1M tokens, works on regulated public administration, industry, and aerospace tasks, and ships under Apache License 2.0.

  2. Why it matters

    Kolibri's efficient design means it can run at lower deployment cost while matching models with up to 4× the active parameters on math, coding, grounding, and long-context tasks, so regulated and mission-critical work can stay on an openly available model without sacrificing capability.

WHO IT HITSEnterprise IT and compliance teams in regulated sectors such as public administration, industry, and aerospace are the clearest audience, as they can deploy a model built specifically for strict compliance and mission-critical use.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, in your inbox every morning.

Context & Analysis

Aleph Alpha released Kolibri on October 3, 2026, timed to German Unity Day, as a sovereign open-weight language model for heavily regulated, mission-critical sectors such as public administration, industry, and aerospace. Kolibri pairs 78 billion total parameters with just 3 billion used per token, and supports English and German up to 1M tokens.

The company frames sovereignty around how a model is built and how it is handed to customers. Kolibri was trained in Germany on German and Finnish infrastructure under European and German law, with an end-to-end pipeline covering data curation, pre-training, post-training, and optimization. Aleph Alpha says its Merlin-Arthur protocol reduces hallucinations by withholding answers when the context does not support them.

Aleph Alpha is also betting on German-language depth: 21.3% of pre-training tokens are German, with minimal translation. The test is whether this transparency and compliance story holds up when regulated buyers compare Kolibri against larger open models under real workloads.

FAQ
What does 'sovereign' mean for Kolibri?
Aleph Alpha defines sovereignty as both how the model is built and how it is transferred to customers. The model is trained in Germany on German and Finnish infrastructure under European and German law, and management control passes to the customer.
How does Kolibri handle German?
Kolibri uses a bilingual German-English tokenizer, with 21.3% of pre-training tokens being German and minimal translation. It is designed from the ground up as a bilingual model rather than an English model that can read German.
What is the Merlin-Arthur protocol?
It is a protocol Aleph Alpha developed to ensure model reliability. When the context does not support an answer, Kolibri withholds it, reducing hallucinations.

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Next articleSyed Anas Mohiuddin finds "protocol pivoting" flaws in AI agents