General-purpose vision-language models have shown strong performance on widely used multimodal benchmarks

yet a systematic understanding of their true capability boundaries in industrial fine-grained perception

visual reasoning

and hallucination resistance remains lacking. In power grid inspection

generic multimodal benchmarks often fail to cover the large variety of domain-specific targets

prominent small-target defects

complex imaging perspectives

and the high cost of false alarms. To address the mismatch between existing generic benchmarks and the demands of real-world power inspection

this paper presents MM-Power

a multimodal benchmark tailored to power grid inspection scenarios. MM-Power contains 2 708 multiple-choice questions organized into three tracks and 12 tasks

covering fine-grained perception

visual-semantic consistency and hallucination assessment

and visual reasoning. The benchmark spans visible-light and infrared modalities

transmission/substation/distribution scenarios

five acquisition perspectives including pan-tilt cameras

unmanned aerial vehicles

robots

handheld cameras

and handheld infrared thermometers

and 130 categories of equipment states

defects

and anomalies. Based on a unified evaluation of 16 representative proprietary and open-weight vision-language models with both English and Chinese prompts

the best model achieves only 77.34% overall accuracy in English and 75.54% in Chinese. Significant weaknesses are observed in existence judgments

negative-sample identification

and spatial relation reasoning

while prompt language also has a noticeable impact on the industrial-scenario performance of some models. These findings indicate that current vision-language models are still far from reliable deployment in practical power inspection

and MM-Power provides a unified benchmark for future evaluation

optimization

and deployment of multimodal models in the power domain.