Aleph Alpha Released Kolibri, an Open German and English Model

So basically: German AI company Aleph Alpha released Kolibri on Saturday, Oct. 3. It is an open-weight model built for German and English. The full weights are on Hugging Face under Apache 2.0, so companies can run it on their own hardware.

The model

It is a mixture-of-experts model. Aleph Alpha’s spec table lists 78.1 billion total parameters and 3.46 billion active per token. It has 384 experts and uses 6 at a time.

Context goes up to 1 million tokens.

It has four reasoning settings: none, low, medium, and high. Aleph Alpha says it supports tool calling and was trained to say it doesn’t know when the documents it’s given don’t hold the answer.

Knowledge cutoff is June 18, 2026.

How it was built

Aleph Alpha says it trained Kolibri on 768 B200 GPUs. Pre-training was 20 trillion tokens over 21 days, followed by more training on longer context.

The blog says German was 21.3% of the pre-training tokens. The company’s press release says around 23% of pre-training data. Those don’t match, and Aleph Alpha didn’t explain the gap.

It says it trained in Germany and Finland, and screened all training data against a blocklist of more than 4.5 million URLs, drawing on sources that include the European Commission’s Piracy Watch List.

The scores

Aleph Alpha ran all the benchmarks with its own setup. On its table, Kolibri scores 96.9 on AIME 2025, 84.3 on GPQA Diamond, and 66.4 on SWE-Bench Verified.

The same table shows Qwen3.8 27B, a dense model, beating Kolibri’s overall English score, 80.2 to 75.5. Aleph Alpha’s pitch is cost per answer, not top score.

What it is not

It is not a frontier model, and Aleph Alpha doesn’t claim it is. It is not a consumer app. The scores are self-reported.

Aleph Alpha also has a pending deal with Cohere to build a joint company. The press release says that deal still needs regulatory approval, and Aleph Alpha runs on its own until it closes.

My take

The interesting part is the license, not the scores. A German company shipping a 78 billion parameter model under Apache 2.0, with a training data writeup aimed at the EU AI Act, gives European buyers an open model they can run themselves and check. It still trails the best open models on its own chart.