This paper investigates the topology optimization of 3D conformal cooling channels. The design is performed directly on the 3D curved surface, without plane-to-surface projection. A within-surface flow (WSF) model is proposed to simulate flow on curved surfaces, especially non-developable surfaces. The WSF model reduces computational cost and structural complexity by simplifying the full 3D design problem into a surface-based one. It operates by confining the flow within a sufficiently thin layer between two frictionless, adiabatic walls. In this work, the thermofluid problems are modeled using a density-based topology optimization method, and three non-developable surfaces—including warped, spherical, and spline surfaces—are selected as case studies. The optimized cooling channel presents a branched layout, and the effects of hyperparameters—including filter radius, fluid energy dissipation threshold, and heat generation coefficient—on the configuration of cooling channels are investigated. The full 3D simulations are constructed based on the topology optimization results, and their performance is compared with reference cooling channels. The validations show the superiority of the topology-optimized channels and highlight the importance of ensuring consistency in inlet Reynolds numbers between the WSF model and full 3D simulations. For topology-optimized cooling channels, the temperature rise reductions compared to reference channels are 6.37% on warped surfaces, 12.89% on spherical surfaces, and 19.38% on spline surfaces. The corresponding pressure drop reductions are 20.94%, 11.58%, and 23.53%, respectively. This work suggests a promising pathway for the design of 3D conformal cooling channels based on topology optimization.