A search for a light pseudoscalar Higgs boson ($\mathrm{a}$) is performed using the 2016 Large Hadron Collider (LHC) proton-proton collision data corresponding to an integrated luminosity of 35.9 fb^-1 collected at √s = 13TeV by the CMS experiment. The search looks for gg ⟶ H and qq ⟶ qqH with H ⟶aa ⟶ μμττ, where H can be the 125 GeV state or a heavier one. Due to the large mass difference between the H and the a and lightness of the a decay products, the two pairs of μ and τ are highly boosted and collimated. A dedicated di-tau reconstruction technique is devised to increase the efficiency with respect to standard tau reconstruction algorithms. In the absence of a significant signal, model-independent limits are set at 95% C.L. on the branching fraction for H ⟶ aa ⟶ μμττ as a function of m_a in the range 3.6 < m_a < 21 GeV$, down to 1.5 x 10^-4}$ for m_H = 125 GeV, 2.0 x 10^-4 for m_H = 300 GeV, and 1.2 x 10^-3 for m_H = 750 GeV. Model-dependent limits are set within the context of two Higgs boson doublet plus singlet models, with the most stringent results obtained for Type-III models. These results are complementary to current LHC limits for heavier m_a that lead to resolved final-state leptons.