We assess the effectiveness of the Jeans Anisotropic Multi-Gaussian Expansion (JAM) technique with a state-of-the-art cosmological hydrodynamic simulation, the Illustris project. We perform JAM modelling on 1413 simulated galaxies with stellar mass M* > 10¹⁰ M_⊙, and construct an axisymmetric dynamical model for each galaxy. Combined with a Markov chain Monte Carlo simulation, we recover the projected root-mean-square velocity (V_rms) field of the stellar component, and investigate constraints on the stellar mass-to-light ratio, M*/L, and the fraction of dark matter f_DM within 2.5 effective radii (R_e). We find that the enclosed total mass within 2.5 R_e is well constrained to within 10 per cent. However, there is a degeneracy between the dark matter and stellar components with correspondingly larger individual errors. The 1σ scatter in the recovered M*/L is 30-40 per cent of the true value. The accuracy of the recovery of M*/L depends on the triaxial shape of a galaxy. There is no significant bias for oblate galaxies, while for prolate galaxies the JAM-recovered stellar mass is on average 18 per cent higher than the input values. We also find that higher image resolutions alleviate the dark matter and stellar mass degeneracy and yield systematically better parameter recovery.