> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Aleph Alpha releases Kolibri, an English-German open-weight model
- URL: https://globalfeed.ai/en/aleph-alpha-releases-kolibri-an-english-german-open-weight-model/
- Published: 2026-10-03T20:09:06.000Z
- Updated: 2026-10-03T20:09:06.000Z
- Description: Aleph Alpha has released Kolibri, an open-weight model for English and German, under the Apache 2.0 license. The model has 78.1 billion parameters in total with 3.46 billion active per token, reads up to 1 million tokens of context and is trained to say it does not know when the context has no answe
- Author: GlobalFeed Editor
- Tags: Aleph Alpha, Kolibri, open-weight models, large language models, x-Aleph__Alpha, dil-en, elle

**Aleph Alpha** released its new language model **Kolibri** on October 3, 2026\. Built by the Heidelberg-based German company, the model is designed for English and German and is available under the Apache 2.0 license. Its weights are open for anyone to download, so organizations can run it on their own servers. The company positions Kolibri as a sovereign model that public bodies and heavily regulated industries can use without sending internal data to an outside service.

## How Kolibri is built

According to the model card, Kolibri uses a Mixture-of-Experts (MoE) architecture. It has 78.1 billion parameters in total, but only 3.46 billion are active for each token. Each layer has 384 experts, and 6 of them are selected at every step. The design aims to bring the knowledge of a large model at the compute load of a much smaller one. The context window reaches up to 1 million tokens; long-context training was done at 256,000 tokens, and the model card recommends that length for efficiency.

Kolibri was trained for English and German only. Aleph Alpha says it devoted a large share of the training data to German text and built its own tokenizer that splits long German compound words more efficiently. The model supports a reasoning mode with four effort levels, none, low, medium and high, so users can trade cost against answer quality. Tool calling and structured output are also supported.

## Trained to say it does not know

According to the announcement, Kolibri is trained to say "I don't know" when the answer is not in the context it was given, instead of making something up. Aleph Alpha uses its own method for this, called the Merlin-Arthur protocol, which generates synthetic negative examples for training and aims to ground the model's answers in their sources. The company stresses that this matters for mission-critical work, where a wrong answer can create compliance risk.

## Who it is for

According to Aleph Alpha, the model targets sectors such as public administration, aerospace, manufacturing, automotive and semiconductors. The model card lists conversational assistants, agentic workflows, document processing, retrieval-augmented generation (RAG) and internal knowledge tools among its uses. The weights are on Hugging Face. According to the model card, running Kolibri requires data center hardware: at least two 80 GB A100 or H100 GPUs, or a single H200, B200 or B300\. It runs on vLLM with the company's aleph-alpha-inference plugin and offers an OpenAI-compatible API. The model card says Kolibri is designed for human-reviewed systems and should be used in an advisory role in decision support.

Kolibri follows Kolibri Origin, released in June. According to the company, in the three months between them the model grew from 30 billion to 78 billion parameters and the context window from 65,000 tokens to 1 million. It was trained in Germany and Finland on NVIDIA B200 GPUs, and its knowledge cutoff is June 18, 2026.

## On the way to merging with Cohere

On September 16, 2026, Aleph Alpha announced a definitive business combination agreement with Canadian AI company Cohere. According to the announcement, the combined company will operate globally as Cohere, with headquarters in Berlin and Toronto, and Aleph Alpha's Heidelberg office will remain a research center. The deal is still subject to regulatory approval, and the companies expect it to close later this year.

The test results in the announcement were shared by Aleph Alpha. Because the model was trained only for English and German, there is no information on how it performs in other languages, and running it takes powerful data center hardware. It is wise to test Kolibri on your own documents and workflows before putting it to use. For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Aleph Alpha: Kolibri announcement](https://aleph-alpha.com/en/blog/kolibri-has-landed-a-sovereign-open-weight-model/?ref=globalfeed.ai), [Hugging Face: Kolibri-1 model card](https://huggingface.co/Aleph-Alpha/Kolibri-1?ref=globalfeed.ai), [Aleph Alpha: Cohere agreement](https://aleph-alpha.com/en/news/cohere-agreement-transatlantic-sovereign-ai/?ref=globalfeed.ai).