Bijan Bowen runs Qwen 3.8 27B locally at high precision and tests it across a browser operating system, several games, a printable engine model, a 3D watch site and more experimental front-end work. The model's compact size and permissive licensing make its practical performance especially relevant to people who want capable local inference on a wider range of hardware.
The results were mixed at first. A browser interface was amusing but included a broken driving camera, and the C++ skate task needed human intervention without becoming reliable. Other work was much stronger: the Subway-style first-person game played well, the engine model showed detailed structure and the watch site combined convincing materials with polished motion and presentation.
The strongest later results included a city timeline and a cinematic game that showed design quality well above what the model's size would normally suggest. The review concludes that Qwen 3.8 27B is one of the most broadly useful local models tested, while leaving open how much lower-precision quantization might reduce that quality on smaller machines.
Watch on YouTube
