Deep earthquakes behave like shallow earthquakes but must have fundamentally different physical processes. Their rupture behaviors, magnitude-frequency statistics, and aftershocks are diverse and imperfectly dependent on various factors, such as slab temperature, depth, and magnitude. The three leading mechanisms for deep earthquakes (i.e., transformational faulting, dehydration embrittlement, and thermal runaway) can each explain portions of the observations but have potentially fundamental difficulties explaining the rest. This situation calls for more serious consideration of hypotheses that involve more than one mechanism. For example, deep earthquakes may initiate by one mechanism, but the ruptures may propagate via another mechanism once triggered. To make further progress, it is critical to evaluate the hypotheses, both single- or dual-mechanism, under conditions as close to those of real slabs as possible to make accurate and specific predictions that are testable using seismic or other geophysical observations. Any new understanding of deep earthquakes promises new constraints on subduction zone structure and dynamics. ▪ Deep earthquakes display the complex structure and dynamics of subduction zones in terms of geometry, stress state, rheology, hydration, and phase changes. ▪ Phase transformation, dehydration, and thermal runaway are the leading mechanisms for deep earthquakes, but all have major gaps or fundamental difficulties. ▪ Deep earthquakes may involve dual-mechanism processes, as hinted at by the diverse rupture and statistic properties and the break of self-similarity. ▪ Further progresses would benefit from specific and testable predictions that consider realistic slab conditions with insights from geodynamics, petrology, and mineral physics.