Location is a key feature of urban spaces. It influences housing and land markets, social segregation, and decisions in urban planning and administration. Hedonic price models and, increasingly, machine learning (ML) methods are traditionally used for quantitative location valuation. At the same time, transformer architectures and their more recent adaptations for tabular data mark a methodological milestone. However, to date, tabular transformers have not been widely adopted in spatial analyses and, in particular, have not been systematically evaluated for location value assessments. This study addresses this gap by comparing classic ML methods with two tabular transformer approaches: TabNet and Mitra. In a case study, rents are used as the target variable, with property-specific characteristics being normalised using spatial regressions. Location-describing features include accessibility to POIs and areal variables such as imperviousness, noise, and population density. While a stacked ensemble of classical ML models achieved the lowest RMSE, TabNet and Mitra performed within a narrow margin, demonstrating that they can be used to map spatial patterns.