The LocalLLaMA poster says Chinese models appear to perform unevenly on new benchmarks, then improve substantially in later releases. They argue this may indicate more benchmark optimization than in American models, while acknowledging that models continue to improve.
The poster attributes the perceived gap partly to differences in data ecosystems and compute, and links a benchmark audit and a post on X. They also describe an image of Tencent internal benchmarks as showing American models ahead, but the article gives no scores or independent confirmation of the image or its interpretation.