Mixing of fluids in a coaxial jet is studied under four distinct viscosity ratios,
$m=1$
,
$10$
,
$20$
and
$40$
, using highly resolved large-eddy simulations (LES), particle image velocimetry and planar laser-induced fluorescence. The accuracy of predictions is tested against data obtained by the simultaneous experimental measurements of velocity and concentration fields. For the highest and lowest viscosity ratios, standard RANS models with unclosed terms pertaining to viscosity variations are employed. We show that the standard Reynolds-averaged Navier–Stokes (RANS) approach with no explicit modelling for variable-viscosity terms is not applicable whereas dynamic LES models provide high-quality agreement with the measurements. To identify the underlying mixing physics and sources of discrepancy in RANS predictions, two distinct mixing modes are defined based on the viscosity ratio. Then, for each mode, the evolution of mixing structures, momentum budget analysis with emphasis on variable-viscosity terms, analysis of the turbulent activity and decay of turbulence are investigated using highly resolved LES data. The mixing dynamics is found to be quite distinct in each mixing mode. Variable viscosity manifests multiple effects that are working against each other. Viscosity gradients induce additional instabilities while increasing overall viscosity decreases the effective Reynolds number leading to laminarization of the turbulent jet, explaining the lack of dispersion and turbulent diffusion. Momentum budget analysis reveals that variable-viscosity terms are significant to be neglected. The scaling of the energy spectrum cascade suggests that in the TLL mode the unsteady laminar shedding is responsible for the eddies observed.