Parking occupancy detection plays a critical role in optimizing urban spaces by enabling dynamic resource allocation and reducing traffic congestion. While prior approaches relying on lightweight architectures offer fast inference, their task-specific designs necessitate architectural modifications for new applications, limiting adaptability. Multimodal Large Language Models (MLLMs) have emerged as versatile alternatives, with efficient variants now deployable on edge devices. This paper presents the first benchmark of an existing efficient MLLM, SmolVLM, for the task of parking occupancy detection. We systematically evaluate the model’s feasibility in a zero-shot setting, along with parameter-efficient fine-tuned variants across different model sizes and user/system prompt configurations, and further assess its ability for cross-dataset generalization. Experiments on two benchmark datasets, PKLot and CNRPark+EXT, demonstrate that our approach exhibits strong in-domain and cross-dataset performance, and either compete with or surpass prior task-specific architectures, despite training on significantly lesser data ( \(91\%\) reduction for PKLot; \(40\%\) reduction for CNRPark+EXT). Source code for replicating our experiments is available at: https://osf.io/kdhe5?view_only=72e17f288d2f471eb4deb464d177a50c