An independent developer tested Nanbeige4.3-33B on 299 Korean questions with the same retrieved documents. Initial reasoning-mode accuracy was 66.2%; rerunning only 29 length-limited cases with a larger output budget raised it to 70.6%, below the tested Qwen3.8-27B result of 77.3%. Different quantization formats and selective rerun conditions prevent attributing the difference solely to model quality or one intervention. The official model card’s coding and agent benchmarks do not validate Korean RAG, so deployments need evaluation on their actual language and documents.