Kolibri-1 is a 78 billion parameter German-English open-weight model that runs efficiently on a single NVIDIA A100 GPU, excelling in reasoning, tool use, and handling extremely long context windows up to one million tokens, making it suitable for organizations prioritizing data sovereignty. While it demonstrates strong capabilities in language tasks and code correction, it faces challenges with high-effort reasoning, visual content generation, and factual accuracy without external references.
A German company recently released a 78 billion parameter open-weight model called Kolibri-1, which operates efficiently by activating only 3.46 billion parameters per token. The model supports German and English languages, features a reasoning mode with adjustable effort levels, and includes tool-calling capabilities. It was trained on European infrastructure under strict data sovereignty laws, making it appealing for organizations like hospitals and banks that require data privacy. The model architecture includes 50 layers with a mix of sliding window and full attention, and uses a mixture of experts approach to balance storage and compute needs, resulting in a smaller cache size compared to dense models.
The video creator tested Kolibri-1 on a single NVIDIA A100 GPU, achieving a generation speed of about 118 tokens per second, which is impressive for a model of this size. However, running the model required patched software support since the official implementation demands high-end hardware and proprietary plugins. The creator converted the model weights into a compatible format and successfully ran it on the GPU, confirming the vendor’s claims about speed and efficiency. Initial tests showed the model could answer German questions accurately and solve complex math problems correctly, demonstrating strong reasoning capabilities.
Despite these strengths, the model struggled with high-effort reasoning tasks, often consuming the entire token budget without producing answers. Visual generation tasks, such as drawing animations or games via code, had mixed results: some worked well, like a snake game, while others, including a Mandelbrot explorer and artistic drawings, failed or were incomplete. The model also showed the ability to self-correct code errors when prompted, highlighting its potential for programming assistance. Tool calling was tested successfully, with the model correctly interpreting and converting natural language inputs into API calls and composing responses based on returned data.
A key feature of Kolibri-1 is its long context window, advertised as supporting up to one million tokens. However, the default configuration caps the context at around 262,000 tokens, requiring manual overrides to unlock the full capacity. The creator tested the model’s ability to recall specific information embedded deep within long texts, successfully retrieving a “needle in a haystack” passphrase at both 196,000 and 719,000 tokens, though generation speed slowed significantly at higher context lengths. The model’s factual knowledge was inconsistent, with some correct answers but errors in detailed trivia, aligning with vendor warnings to provide documents for reference rather than relying on memorized facts.
In conclusion, Kolibri-1 is a promising open-weight large language model for German and English that excels in reasoning, tool use, and handling long documents while running efficiently on a single high-memory GPU. However, it has limitations in high-effort reasoning tasks, visual content generation, and factual accuracy in closed-book scenarios. Its large context window is a standout feature but requires configuration tweaks to fully utilize. The model is best suited for organizations needing data sovereignty and long-context processing rather than for coding agents or fact-recall applications at this stage. The video provides detailed insights into running and testing the model, with links to conversion tools and further resources for interested users.