Aleph Alpha has landed a new model, and it is named Kolibri. The company released the English-German Mixture-of-Experts Transformer on the Day of German Reunification, with 78B total parameters, 3B active, and a context length of up to 1M tokens. The model is downloadable with full weights on Hugging Face under the Apache 2.0 license.
The release comes with a detailed tech report explaining how Kolibri was built. Aleph Alpha started with Kolibri Origin, a 30B total, 3B active model with a 65k token window, and ran the newer model through the same pipeline: data ingestion, curation, ablations, pre-training, post-training, and final evals. The company says the pipeline allowed it to run hundreds of ablation experiments and keep training stable even when hardware failed or data connections dropped.
What Kolibri Does
Kolibri is built for sovereign mission-critical work in regulated areas. That includes public administration, industrials, and aerospace. The model is specialized for German, reasoning, math, agentic behavior, and other capabilities customers need in production. The aim is to optimize performance for each customer’s specific use case, so they can monitor the economic impact and measure ROI over time.
Sovereignty is a core promise. Aleph Alpha says it combines two dimensions: how the model was built and how it transfers to customers. The company offers full supply-chain integrity, accounting for every decision from data ingestion through pre- and post-training to final evals. Customers get full deployment freedom and IP safety, with compliance inherited as a property of the model.
The model is available for download with full weights on Hugging Face under the Apache 2.0 license. The company says the time invested in building and iterating on the pipeline was a valuable investment, judging by how much better Kolibri is than Kolibri Origin and how little time separates their releases.
Comparing the Capabilities
Kolibri sits on the Pareto frontier for quality versus serving cost, using 3B active parameters out of 78B total. The company compares it against other models and finds none deliver more quality at the same serving cost or the same quality at lower cost.
| Capability | Kolibri | Nemotron 3 Super |
|---|---|---|
| Active params | 3B | Up to 4x Kolibri |
| Total params | 78B | Not specified |
| Context length | 1M tokens | Not specified |
The comparison shows Kolibri’s efficiency. It matches models with up to four times its active parameter count while holding the serving cost down.
Training Details
Kolibri was trained with abstention data and the Merlin-Arthur protocol. That means it learns to say “I don’t know” when the answer is not in the context. The company tracks and validates abstention accuracy continuously.
The model is bilingual by design, not an English model that has read some German. Aleph Alpha developed a German/English tokenizer and included organic German data throughout training, so that 21.3% of the pre-training tokens are German. Translation was used sparingly, at 6% overall, since translated text carries the cultural fingerprint of its source language.
Kolibri was built with the EU AI Act, the General-Purpose AI Code of Practice, and the GDPR in mind from the start. Copyright law was a focus of the work toward trustworthy technology. The company is transparent about model weights and training data curation, so decisions behind development are visible.
Reasoning traces make the model explainable. With the Merlin-Arthur protocol, the model grounds trustworthiness by refraining from answering when the context doesn’t support an answer. The model was built in Germany, trained on infrastructure in Germany and Finland, under European and German law, with no foreign control. Aleph Alpha owns the entire pipeline, from data curation through pre- and post-training to optimization in its Model Factory.
The Bottom Line
Kolibri is a specialized model for regulated industries, built with full transparency and sovereign deployment in mind. The 1M-token context length and 3B active parameters make it usable on-premise without sending internal data to third-party inference services.
The comparison to Nemotron 3 Super shows the model’s efficiency. Kolibri matches models with up to four times its active parameter count while holding the serving cost down.
The release is notable because it brings together several strands at once:
- A large, specialized model with a broad context window
- Full weights downloadable on Hugging Face under Apache 2.0
- Detailed documentation of the training pipeline
- Sovereignty claims backed by full supply-chain integrity
The question now is whether customers find the specialized capabilities and sovereignty guarantees enough to justify moving to Kolibri. The company has made its case public.
Source material: “Kolibri Has Landed: A Sovereign Open-Weight Model,” aleph-alpha.com.
Get the Notebook.
The day's best stories and every fresh verdict, in plain English, in your inbox by seven. One email a day, no more.