Aleph Alpha releases Kolibri, an open-weight model that makes the case for European AI sovereignty
Aleph Alpha released Kolibri, a 78-billion-parameter open-weight German-English mixture-of-experts model.
Aleph Alpha released Kolibri, an Apache 2.0 open-weight German-English model with 78 billion parameters and about 3 billion active per token through a mixture-of-experts design. German accounts for 21.3 percent of the training data via a dedicated pipeline, and Chinese models were used to generate some synthetic data. Aleph Alpha says Kolibri scores 71 percent on German benchmarks and decodes faster than GPT OSS A5B, Qwen 3.6 A3B, and Gemma 4 A4B. It targets public administration, aviation, and industry, supports up to a one-million-token context window, and was trained on 768 NVIDIA B200 GPUs in Germany and Finland.
- Kolibri is a 78B-parameter MoE model with about 3B active parameters.
- Apache 2.0 weights are on Hugging Face; German is 21.3% of data.
- Aleph Alpha reports 71% on German benchmarks and faster decoding than peers.
- Training used 768 B200 GPUs in Germany and Finland.
- Context window reaches one million tokens for government and industry use.
Full article227 words · extracted from the-decoder.com · click to collapse
Oct 5, 2026
Aleph Alpha has released Kolibri, a German-English language model with 78 billion parameters, about three billion of which are active per token through a mixture-of-experts architecture. Aleph Alpha claims Kolibri sits on the Pareto front of quality and operating cost in both languages, outperforming compared models with similar architectures, some significantly older (March, April 2026), on either metric. German accounts for 21.3 percent of the training data, backed by a dedicated German data pipeline the company built for the project. Chinese models were also used to generate synthetic training data.
Aleph Alpha says Kolibri targets public administration, aviation, and industry. The model was developed under European law with the EU AI Act in mind, supports context windows of up to one million tokens, and was trained on 768 B200 GPUs in Germany and Finland, according to the tech report. The weights are available under an Apache 2.0 license on Hugging Face.
AI News Without the Hype – Curated by Humans
Subscribe to THE DECODER for ad-free reading, a weekly AI newsletter, our exclusive "AI Radar" frontier report six times a year, full archive access, and access to our comment section.
Text extracted automatically; images, tables and formatting may be missing. Original: https://the-decoder.com/aleph-alpha-releases-kolibri-an-open-weight-model-that-makes-the-case-for-european-ai-sovereignty/