This work investigates a hierarchical Siamese multi-task learning approach for vehicle model similarity, using the StanfordCars dataset as a benchmark. The main contribution is the combined learning objective, which aims to jointly learn to compute the similarity between two, or more hierarchic concepts (i.e., car make comparison and the fine-grained car model similarity), a novel technique, especially for vehicle model classification. Two multi-task architectures are presented and evaluated: parallel and cascaded, which incorporate auxiliary objectives such as car make and car type similarity to enhance the main task of model-level similarity estimation. The experiments span nine established classifiers and examine both single-task and multi-task configurations. The results demonstrate that multi-task learning improves performance by up to 8.8% over single-task baselines. Notably, four of the top five configurations leverage a multi-task setup, with the Parallel EfficientNet model, and make similarity objective achieving the highest accuracy (86.9%). Incorporating car make as an auxiliary objective consistently boosts model similarity accuracy by 1–2%, and parallel architectures outperform cascade in most scenarios. Given the complexity of vehicle make and model classification (VMMC), a domain with hundreds of thousands of class variants and rapid model turnover, the findings support multi-task learning as a scalable solution. This is particularly relevant for intelligent transportation systems, where accurate recognition of vehicle types is essential for applications in traffic monitoring, driver assistance, and smart city applications.