Aleph Alpha Kolibri: What a Sovereign Open-Weight AI Model Means for Your Business

Where your AI model runs, and who controls it, is becoming a board-level question. On October 3, Aleph Alpha gave that question a new option: a large open-weight model you can download and run yourself.

Aleph Alpha released Kolibri, an English-German model with the full weights on Hugging Face under the Apache 2.0 license. This post covers what the release contains, how to read the benchmark claims, and what it means if you are planning an AI digital transformation in a regulated industry.

What Is Aleph Alpha Kolibri?

Per Aleph Alpha’s announcement, Kolibri is a Mixture-of-Experts transformer with 78 billion total parameters, of which about 3 billion are active for each token. It supports context lengths of up to 1 million tokens. The company says it is built for sovereign, mission-critical work in regulated areas such as public administration, industrials and aerospace.

An independent write-up by Traictory checked the model repository and confirmed the provable parts: the 78B parameter count, the Apache 2.0 license and the million-token context window. It puts the active size at roughly 3.46 billion parameters per token, which Aleph Alpha rounds down to 3 billion.

What Does “Sovereign” Mean Here?

Aleph Alpha describes sovereignty as two things: how the model was built, and how it transfers to the customer. The company says it accounts for every step from data ingestion to final evaluation, and that customers get freedom of deployment and intellectual-property safety. Because the model is small for its quality, Aleph Alpha says it can run on-premise without sending internal data to third-party inference services.

In plain terms, sovereign AI usually means three things for a business:

• Your data stays where you decide. The model runs in your environment, not in someone else’s.
• You are not tied to one vendor. An open-weight model with a permissive license can be moved or replaced.
• You can show regulators how it works. Documented training and deployment make compliance reviews easier.

What Do the Benchmarks Show?

Aleph Alpha says Kolibri matches models with up to four times its active parameter count, such as Nemotron 3 Super, on math, coding, grounding and long-context tasks. Treat this carefully. Traictory points out that every score in Aleph Alpha’s comparison table is the vendor’s own, run by Aleph Alpha and not by the labs that built the other models.

Traictory’s reading of the table is also useful. Kolibri leads on the AIME math rows and on GPQA diamond in both English and German. It trails Qwen3.6-35B on several tool-use tests, including tau2-bench retail, tau2-bench telecom and BFCL v4. For a business building AI agents that call tools, that second group of results matters more than the math scores.

Aleph Alpha also reports internal tests built to mimic customer work in sectors such as the German public sector, manufacturing and aerospace. These are the vendor’s own test suites, so they show direction, not proof. The only test that counts is one on your own data and workflows.

Should You Switch to an Open-Weight Model?

Not automatically. This is our view, not Aleph Alpha’s. Open weights give you control, but you take on the work of hosting, monitoring and updating the model. A hosted model from a large provider can still be the better choice for a first pilot.

Open-weight models make the most sense when:

• You handle data that cannot leave your own infrastructure or region.
• Your sector requires you to document and audit how AI decisions are made.
• You want a fallback so one provider’s pricing or access change cannot stop a workflow.
• Your users work in German or English and you want to test a model tuned for both.

How to Plan Without Betting on One Model

Models change every few weeks. Build so that you can change with them:

1. Pick one workflow. Choose a process with clear inputs and outputs, such as document review or support triage.
2. Write a test set from your own data. Use real examples to compare two or three models, including one open-weight option.
3. Keep the model behind a thin layer. Your applications should call your own interface, so switching models does not mean rewriting them.
4. Decide where it runs. Hosted, private cloud or on-premise, based on the data involved.
5. Log every action. Record what the model was asked and what it did.

How Incresco Helps

Incresco does AI digital transformation. Our AI transformation work covers the pieces above:

• AI strategy and consulting: roadmaps, readiness assessments and the business case, so you pick the right process and the right kind of model first.
• Custom AI development: AI agents, copilots and intelligent automation built around your workflows.
• AI integration: connecting AI to your existing systems through workflow automation, API integration and legacy system modernisation.

If you want to compare hosted and open-weight models on one of your own workflows, talk to our team.