“The highest-scoring open-weight model, Qwen-VL-8B-Instruct, achieved an 81% accuracy score on the ChinaHeritaQA benchmark, outperforming the average human accuracy of approximately 67%.”