Germany's NEW AI Model is INSANE (FREE!)

Colibri: Germany's New Open-Weight AI Model

Aleph Alpha, a German AI company, released Colibri on October 3rd—a 78-billion-parameter mixture-of-experts model that is free to download under the Apache 2.0 license. Designed for speed and efficiency, Colibri outperforms models like Qwen, Nemotron, and Mistral on several benchmarks, marking a significant entry from Europe into the open-weight AI space.

Model Architecture and Capabilities

Colibri uses a mixture-of-experts (MoE) architecture: while the full model has 78B parameters, only about 3B are active at any given moment. This allows it to remain fast despite its size. It also features a context window of up to 1 million tokens, enabling it to process very long documents or conversations without losing track of earlier context.

The model was trained on 24 trillion tokens using a hybrid attention mechanism. Notably, German makes up over 20% of its training data, supplemented by proprietary German-language datasets. As a result, Colibri is optimized for German and English—other languages are not supported.

Licensing and Local Deployment

Colibri is open weights under the Apache 2.0 license, meaning the model files are publicly available on Hugging Face. Users can download and run it on their own machines without any account, sign-up, or monthly fee. This also ensures data privacy: all processing stays local, making it ideal for sensitive client information or business ideas.

However, there is a critical hardware requirement: even though only 3B parameters are active, the entire 78B model must be loaded into memory. A standard laptop may struggle; users should check their computer’s RAM against the file size on Hugging Face before downloading. Smaller quantized versions may become available later.

Benchmark Performance

Aleph Alpha released results from two key benchmarks:

• German benchmark (average across multiple German tests): Colibri scored highest, beating Nemotron, Gemma 4, Qwen 3.6, and GPT-Aust. This is expected given its German-language focus.
• Tool-use/agent benchmark: Colibri performed well against Mistral Small 4, Qwen 3.6, and Nemotron 3 Super (Nvidia), with Mistral ranking lowest.

The video notes that benchmarks are selective—companies often highlight favorable tests. The real test is how the model performs on individual tasks, especially for agentic tool calling (e.g., using tools and completing multi-step tasks). Colibri is described as particularly suited for such agent workflows.

Practical Considerations

• Text-only: Colibri does not generate images or video.
• Language limitation: Only German and English; not suitable for multilingual use.
• Hardware: Requires sufficient RAM to hold the full 78B model. Use LM Studio or similar tools for easy local setup.
• Recommendation: Test the model yourself rather than relying solely on benchmarks. Early adopters who experiment with local models often gain an edge.

Key Takeaway

Colibri demonstrates that a small, fast, open-weight model from Europe can compete with major players—and it is freely available for anyone to run locally. Its strong German-language performance and agent-friendly design make it a promising option for privacy-conscious users and businesses. The main barrier is hardware; check your system’s memory before downloading. If you work with German or English and value data sovereignty, Colibri is worth a try.