Kimi open-sourced PerceptionBench, a visual perception benchmark with 3,000 manually verified questions. Evaluation results show GPT-5.6-Sol led 16 multimodal large models with 59.7% accuracy, while no model exceeded 60%.