I Ran Germany's New #1 Open Source AI Model... Kolibri 1 🇩🇪

Summary

The video reviews Kolibri 1 (also transcribed in the subtitles as “Calibri 1”), describing it as a German-developed, open-weight model for German and English. The narrator says it is a mixture-of-experts model with 78 billion total parameters and about 3 billion active parameters, released under the Apache 2.0 license. It offers low, medium, and high reasoning modes.

Benchmarks and performance

• The narrator cites a score of 75 on one English evaluation and an 88.9 average on SWE, claiming strong results compared with models such as Qwen 3.5 35B, GLM-4.5 Air, Mistral, and GPT-OSS 120B. These are reported benchmark results; the video’s hands-on coding demonstrations are less impressive.
• On the presenter’s system, generation reached roughly 61 tokens per second without speculative decoding. The narrator also reports a run using about 77 GB of memory.
• Reasoning mode sometimes improved performance on math and logic tasks, but did not consistently fix errors. The model also sometimes looped or became stuck, especially during extended reasoning.

Tests and demonstrations

• Math and logic: The model solved one math query without reasoning and an IMO-style problem with low-level reasoning. It handled the oranges-and-children puzzle but failed other familiar reasoning questions, including the surgeon riddle.
• Software and platform knowledge: It gave a partly correct iOS/Swift answer but struggled with C++ vectors and gave an incorrect answer about Windows Store/WinRT publishing.
• Culture and trivia: It missed references to British comedy, music, and pop culture, suggesting gaps in English-language cultural knowledge.
• Code generation and visual demos: The presenter tried browser and Python projects, including a solar system, a volcano, face and anatomy visualizations, and simple games. Results ranged from basic or appealing demos to syntax errors, runtime errors, looping, and incomplete output. German comments occasionally appeared in code prompted in English.
• App recreations: A Word-like interface supported basic formatting, a Photoshop-like demo produced a smiley face, and a Blender-like demo allowed the presenter to manipulate and recolor a cube. PowerPoint and iMovie attempts encountered errors or were incomplete.

Overall assessment

The presenter considers Kolibri 1 fast and potentially useful for simple application logic and lightweight demos, but not dependable for substantial software engineering, complex interactive projects, or medical and anatomy work. The review emphasizes that this is an early version and that results may differ when prompting in German.

The narrator also explains temperature as randomness in token selection: higher temperature can produce less predictable outputs and may help when the model gets stuck, while zero temperature is presented as more deterministic. The examples show that changing temperature can affect whether generation progresses, but does not guarantee correct code.

Main speaker and sources

• Main speaker: The video’s presenter and narrator from the xCreate channel.
• Other sources discussed: Kolibri 1’s cited benchmark results and the presenter’s own prompts and generated outputs.