A Reddit user reports that Qwen3.8-27B has significantly degraded in performance, failing at basic tasks and forgetting instructions despite previously excelling at complex coding and bug fixes. The author suspects the model was secretly quantized or nerfed to make the upcoming Qwen4 appear superior, also noting increased "laziness" in task completion. The issues emerged when running a Q4 quantization via llama.cpp on an RTX 4090.