GitHub radar
Kolibri: German-English AI model for reasoning, tools, and code
Aleph Alpha released Kolibri, an open mixture-of-experts model built for step-by-step reasoning, tool calling, and agentic work in German and English, including code. You can only run it on a serious server.
It's a large open AI model that reasons step by step, calls outside tools and programs on its own, and handles German and English text, including code.
When it helps
Useful if you need an advisor for complex reasoning, agentic workflows, or code in German and English, and you have access to a powerful server. Not useful if you want to run something on a regular computer — it's too heavy for that.
Pros
- Open weights under Apache 2.0 — free to use and modify
- Holds over a million tokens of text in memory at once
- Reasons step by step and calls outside tools on its own
- Developers' own benchmarks show solid code results, including SWE-Bench
Cons
- Can't run at home — needs at least two A100 80GB server GPUs or one H200
- Developers recommend keeping it as an advisor, not letting it decide on its own
- Built mainly for German and English, other languages aren't the focus
How to set it up — step by step
- 1Check whether you have access to a server with two A100 80GB GPUs (or one H200/B200).
- 2Open your AI agent (Claude Code or Codex) in your project folder.
- 3Send the agent the text from the block below.
- 4Check how the model answers in German and English with reasoning and tool calling turned on.
- 5Judge the quality of its answers and code for your task before trusting it with final decisions.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
If you have access to a server with the required GPUs, deploy Aleph Alpha's Kolibri-1 model from https://huggingface.co/Aleph-Alpha/Kolibri-1 via vLLM, ask it something in both German and English, and show me how the reasoning mode and tool calling work.
This is a server-class model — you cannot run it at home. It needs at minimum two A100 80 GB accelerators (or one H200/B200) — tens of gigabytes of GPU memory.
Open on Hugging Face▌ More finds