Thermal evolution models are presented for Ganymede, assuming a mostly differentiated initial state of a water ocean overlying a rock layer. The only heat sources are assumed to be primordial heat (provided by accretion) and the long-lived radiogenic heat sources in the rock component. As Ganymede cools, the ocean thins, and two ice layers develop, one above composed of ice I, and the other below composed of high-pressure polymorphs of ice. Subsolidus convection proceeds separately in each ice layer, its transport of heat calculated using a simple parameterized convection scheme and the most recent data on ice rheology. The model requires that the average entropy of the deep ice layer exceeds that of the ice I layer. If the residual ocean separating these layers becomes thin enough, then a Rayleigh-Taylor-like (“diapiric”) instability may ensue, driven by the greater entropy of the deeper ice and merging the two ice mantles into a single convective layer. This instability is not predicted by linear analysis but occurs for plausible finite amplitude perturbations associated with large Rayleigh number convection. The resulting warm ice diapirs may lead to a dramatic “heat pulse” at the surface and to fracturing of the lithosphere, and may be directly or indirectly responsible for resurfacing and grooved terrain formation on Ganymede. The timing of this event depends rather sensitively on poorly known rheological parameters, but could be consistent with chronologies deduced from estimated cratering rates. Irrespective of the occurrence or importance of the heat pulse, we find that lithospheric fracturing requires rapid stress loading (on a time scale ⪅10⁴ years). Such a time scale can be realized by warm ice diapirism, but not directly by gradual global expansion. In the absence of any quantitative and self-consistent model for the resurfacing of Ganymede by liquid water, we favor resurfacing by warm ice flows, which we demonstrate to be physically possible, a plausible consequence of our models, compatible with existing observations, and a hypothesis testable by Galileo. We discuss core formation as an alternative driver for resurfacing, and conclude that it is less attractive. We also consider anew the puzzle of why Callisto differs so greatly from Ganymede, offering several possible explanations. The models presented do not provide a compelling explanation for all aspects of Ganymedean geological evolution, since we have identified several potential problems, most notably the apparently extended period of grooved terrain formation (several hundred million years), which is difficult to reconcile with the heat pulse phenomenon.