Kolibri: A Sovereign Open-Weight Model

Released on the anniversary of German Reunification (October 3, 2026), Aleph Alpha has officially unveiled Kolibri, a new frontier in sovereign AI.

🛠️ Technical Architecture

Kolibri is engineered as an English-German Mixture-of-Experts (MoE) Transformer. Its architecture is designed for high efficiency, balancing a large knowledge base with lean execution.

---

🚀 The Evolution of the Training Pipeline

Kolibri is not a standalone effort but the result of a rigorous, iterative development cycle. The team first established a robust training infrastructure by creating Kolibri Origin (a 30B total / 3B active model with a limited $\approx 65\text{k}$ token context window).

The pipeline was designed to be manual and fragile fully automated and resilient, ensuring that hardware failures or connectivity drops did not interrupt the process.

The Development Workflow

By investing heavily in this pipeline, Aleph Alpha significantly reduced the time between the release of Kolibri Origin and the current Kolibri model, while drastically increasing overall performance.

---

🛡️ Defining "Sovereignty"

For Aleph Alpha, sovereignty is not a buzzword but a dual-dimensional framework:

"Sovereignty combines two dimensions: how we built the model, and how it transfers to our customers."

The Sovereignty Checklist

• Supply-Chain Integrity: Total transparency and accountability for every stage, from initial data ingestion to final evaluation.
• Deployment Autonomy: Customers maintain full control over where the model is hosted and ensure their intellectual property remains secure.
• Inherited Compliance: Because of the model's structure and deployment freedom, regulatory compliance is a built-in feature.

Kolibri is specifically tailored for mission-critical operations in highly regulated sectors, such as:

• 🏛️ Public Administration
• 🏭 Industrial Manufacturing
• 🚀 Aerospace

---

📈 Performance and the Pareto Frontier

Kolibri is designed to optimize the trade-off between intelligence and operational overhead. In economic terms, it sits on the Pareto Frontier.

$\text{Pareto Frontier} \approx \max(\text{Quality}) \text{ given } \min(\text{Serving Cost})$

This means no other compared model provides higher quality at the same cost, nor the same quality at a lower cost. Because it only uses 3B active parameters, it can be run on-premise, eliminating the need to send sensitive data to external inference providers.

Foundational Benchmarks

Kolibri competes with models possessing up to four times its active parameter count (e.g., Nemotron 3 Super).

Performance Comparison Table (Scores on a 0–100 scale; higher is better)

Summary of Capabilities

Kolibri acts as a balanced generalist, providing competitive results across several key domains:

1. Mathematics & Coding: High proficiency in AIME and HumanEval+.
2. Knowledge: Strong performance in GPQA (both English and German).
3. Agentic Behavior: Specialized for banking, retail, and telecom workflows.
4. Long Context: Capable of handling extensive documents via AA-LCR.

Figure 1: Performance vs. throughput for post-trained models in English and German.

While public benchmarks provide a baseline, Aleph Alpha continues to develop internal evaluation suites to better reflect the nuanced requirements of their professional vertical customers.