Aleph Alpha Releases Kolibri, an Open Weight German and English Model Under Apache 2.0

German AI company Aleph Alpha released Kolibri on October 3, 2026, an open weight language model for German and English with 78 billion total parameters, about 3 billion of them active per token. The weights are on Hugging Face under the Apache 2.0 license, and the model supports a context window of up to 1 million tokens, the company said.

Aleph Alpha released the model on the Day of German Reunification and framed it as a sovereign model that organizations can run on their own premises without sending internal data to outside inference services. It drew wider attention on October 5, when The Decoder and the TLDR AI newsletter covered it and it was discussed on Reddit's r/LocalLLaMA. Because the full weights are public, it belongs to our open source coverage.

What is Kolibri?

Kolibri is a mixture of experts model, which means only a small set of its internal "experts" runs for each token, so it computes like a much smaller model while storing far more knowledge. Aleph Alpha says it has 78 billion parameters in total and about 3 billion active per token, and it supports a reasoning mode and tool calling.

The model is bilingual by design, according to the company. German accounts for 21.3% of the pre-training tokens, about 4.3 trillion tokens, and translated text makes up only 6% of the data overall. Aleph Alpha also built a separate German and English tokenizer, the component that splits text into pieces the model can read.

Aleph Alpha says it trained Kolibri with abstention data so that it answers "I don't know" when the answer is not in the documents it was given. That matters for its target customers: public administration, industry and aerospace.

How was it trained?

Aleph Alpha trained Kolibri on 768 Nvidia B200 GPUs on infrastructure in Germany and Finland, in three stages totaling nearly 24 trillion tokens, according to its technical write up. Pre-training on 20 trillion tokens took 21 days.

The company says it first validated its pipeline with a smaller model, Kolibri Origin, with 30 billion total parameters and a 65,000 token context window, then scaled to Kolibri. For supervised fine tuning it generated 174 billion tokens of synthetic data, and reinforcement learning ran on more than 1.2 million curated tasks, Aleph Alpha says. The knowledge cutoff is June 18, 2026, according to the model card.

How does it compare with other open models?

On Aleph Alpha's own benchmarks, Kolibri leads several similar sized open models on math and some agent tests and trails on others. The scores below are reported by Aleph Alpha and have not been independently reproduced.

Aleph Alpha also says Kolibri sits on the "Pareto frontier" of quality versus serving cost in both English and German, meaning no compared model delivers more quality for the same throughput. The Decoder notes that some of the models Aleph Alpha compared against were released in March and April 2026.

Who can run it, and on what?

Anyone can download Kolibri from Hugging Face, and it runs on a single high end GPU. The model card lists a footprint of about 78 GB with FP8 weights, an 8 bit number format that halves memory use, and a minimum of two A100 80 GB or two H100 cards, or one H200, B200 or B300.

Aleph Alpha provides a plugin for vLLM, a popular open source serving engine, and a ready made container image. The company says it built the model with the EU AI Act, the General Purpose AI Code of Practice and the GDPR in mind.

What we don't know yet

• How Kolibri performs in independent evaluations, in German and in English.
• Whether Aleph Alpha will publish the training data or only its curation process.
• What hosted API pricing, if any, Aleph Alpha will offer alongside the free weights.