Aleph Alpha Releases Open-Weight Kolibri-1 Model for German and English

The German company Aleph Alpha has released Kolibri-1, a language model with open weights under the Apache 2.0 license. The license lets companies use, run, and modify it freely. It is a reasoning model. It has an explicit step-by-step mode for working through tasks, and it can call external tools. The model is built for agentic tasks, for writing code, and for retrieval-augmented generation, where it pulls in outside documents before answering. It works in German and English. It can take in up to 1,048,576 tokens of text at once. It has 78 billion parameters in total, but only 3.46 billion activate for any given answer, while the rest sit idle. To use it, a company runs the model behind an OpenAI-compatible server. That means the same request format as the OpenAI API. The company then sends it chat requests, the same way code already talks to other AI models.

For European companies, this is an open option they can deploy on their own servers instead of sending data to an outside provider. The license allows that. Running a 78-billion-parameter model still takes serious hardware, though. You need at least two Nvidia A100 or H100 cards, or a single H200, B200, or B300 card. A home computer cannot handle it. In agentic coding, Anthropic's Opus and Fable models are still stronger. Kolibri-1 is not competing with them there.