Early analyses of exoplanet statistics from the Kepler Mission revealed that a single population of multiple-planet systems with low mutual inclinations (1-2 deg) adequately describes the multiple-transiting systems but underpredicts the number of single-transiting systems. Ten years later, the explanation of this so-called “Kepler dichotomy” remains uncertain. The leading hypothesis is that there are at least two (but perhaps a continuum) of sub-populations of intrinsically multi-planet systems with different mutual inclination dispersions. However, the statistical properties and physical origins of these sub-populations are still poorly constrained. In this work, we derive constraints on the intrinsic mutual inclination distribution by statistically exploiting Transit Duration Variations (TDVs) of the Kepler planet population. Planet-planet interactions generate orbital precession that leads to slow drifts of a planet’s transit duration. These TDV signals are inclination-sensitive and detectable for nearly two dozen Kepler planets. We consider simulated Kepler planet populations from two empirically-calibrated forward modeling frameworks with different assumptions for the mutual inclination distributions. We compute the TDV statistics (specifically, the frequency of detectable TDV signals) of the simulated planet populations and compare them to TDVs of the observed Kepler planets. We find strong evidence for a non-dichotomous, broad, multiplicity-dependent distribution of mutual inclinations. These results place valuable constraints on the physical mechanisms responsible for generating the Kepler planet mutual inclinations.