Kolibri, an open-weight sovereign model from Aleph Alpha with 78 billion parameters and 3 billion active for local use

German company Aleph Alpha announced on 3 October 2026 the launch of Kolibri, an open-weight language model designed for sovereign use in regulated sectors such as public administration, industry and space. The model is released under the Apache 2.0 license with full weights on Hugging Face, and arrives on German Unity Day as a symbol of European technical sovereignty.

Mixture-of-Experts Engineering for Efficiency and Sovereignty

Kolibri is a bilingual Mixture-of-Experts model (English-German) with a total of 78 billion parameters, while the active parameters per token remain at 3 billion. The model supports context lengths up to one million tokens, making it suitable for long documents and deep analysis. This architecture places the model on the Pareto frontier of quality versus service cost: no other model achieves higher quality at the same service cost, or the same quality at lower cost, for both English and German. In general benchmarks, Kolibri matches models whose active parameters are up to four times larger, such as Nemotron 3 Super (120 billion active parameters).

Vertical Specialisation as an Alternative to General Benchmarks

Aleph Alpha recognises that general benchmarks fail to capture the needs of specialised sectors, so it developed internal evaluation suites for the sectors that matter to its customers: the German public sector, aviation, manufacturing and the automotive industry. Each suite reflects the skills, workflows and edge cases required in the sector, and coupled synthetic training environments allowed the model to be improved without ever training on customer data. The results show a consistent improvement across all five sectors from one checkpoint to the next: automotive supplier (0.72 to 0.99), semiconductors (0.35 to 0.80), German public sector (0.54 to 0.75), industrial engine technology (0.31 to 0.60), space (0.14 to 0.59).

Referral and Abstention: Designing Reliability

Kolibri was trained on abstention data and the Merlin-Arthur protocol, which makes it abstain from answering when the context does not support a response. Customers request this feature and its usage is continuously tracked and its accuracy verified. In addition, the model displays its reasoning traces to explain how it arrived at a particular answer, supporting the explainability required in regulated environments.

German by Design, Not by Translation

A bilingual German-English tokenizer was developed, and the team focused on incorporating organic German data throughout training, resulting in German tokens accounting for 21.3 % of the pre-training data. Translation was used sparingly (only 6 %) because translated text tends to carry the cultural imprint of its source language. The result is a model that is bilingual by design, not an English model that read some German.

Sovereignty: Transparent Supply Chain and Free Deployment

Aleph Alpha defines sovereignty in two dimensions: how the model is built and how it is delivered to customers. The company provides full supply-chain safety and accountability for every decision, from data ingestion through pre-training and post-training to final evaluations. Transparency is granted, and customers enjoy unrestricted deployment and intellectual-property security, making compliance an inherited property of the model. The model was built in Germany and trained on infrastructure in Germany and Finland, under European and German law, without foreign control. The company owns the entire production line, from data curation through pre-training and post-training to optimisation in the Model Factory. This end-to-end control establishes model sovereignty.

Model Factory: From Raw Data to Model in Three Months

Between Kolibri Origin (30 billion parameters, 65 k token context, 7.5 trillion training tokens) and Kolibri (78 billion parameters, context up to one million tokens, 20 trillion training tokens), only three months elapsed (June to September 2026). The Model Factory turned model development into exported code: every proposed change launches a small end-to-end experimental run (training, evaluation) to expose regressions within minutes. Trainings are executed as GitHub Actions, and reproducibility means checking out a commit. Checkpoints drop roughly every hour and are automatically evaluated on English and German knowledge, mathematics, code, instruction following, tool use, long context, safety, abstention from hallucination, and grounding.

During the first 21 days of Kolibri’s pre-training, the system recorded 38 unplanned interruptions (about one per 10 000 GPU hours), which were handled automatically without human intervention: the cluster restarted the tasks on different nodes, and training resumed from a checkpoint at most 250 steps away.

What This Means for Readers in the Gulf, Egypt and the Levant

Releasing Kolibri under an open licence with full weights, with an explicit focus on sovereignty, compliance with European law (EU AI Act, GDPR, the General AI Code of Practice), and the ability to deploy locally without sending internal data to external inference services, establishes a reference model for “sovereign capability” rather than merely “sovereign infrastructure”. For organisations in the region that seek models that can run inside their own environments, in multiple languages, with inherited auditability and compliance, Kolibri shows that the technical path exists: small active MoE models, long context, vertical specialisation, and design-level sovereignty. The remaining challenge is not model availability but building an operating model that unites architecture, data, security, governance and skills, which is exactly what we discussed in today’s parallel Saudi story.