A proposal to localising urban AI: a conceptual shift from generalist LLMs to task-specific SLMs

Urban AI discourse has been dominated by Large Language Models (LLMs), yet these models misalign with the specific operational needs of cities, which demand low-latency, context-sensitive, and infrastructure-light solutions. This opinion paper addresses that gap by proposing Small Language Models (SLMs) as a viable alternative and introduces “SLM Urbanism,” a layered conceptual framework that reimagines urban AI deployment. Drawing on recent literature, the framework comprises five layers, computational, task-specialised, application, governance, and citizen-centric, each aligning technical affordances with urban imperatives. Using a normative, design-oriented method, the study contrasts SLMs’ low-cost, edge-native, and interpretable architectures with the compute-heavy, opaque nature of LLMs. The discussion situates SLMs as enablers of locally tuned, explainable, and democratically aligned intelligence that better serve urban equity and efficiency goals. Findings highlight that SLMs often outperform LLMs in resource-constrained settings, enhancing trust, transparency, and civic agency in AI-mediated governance. Importantly, the paper does not reject LLMs entirely but advocates a hybrid future of modular urban intelligence where SLMs lead a shift from centralised automation to distributed, planner-guided agency.