/ AbstractWe present an improved zodiacal light (ZL) model for COBE-DIRBE derived through global Bayesian analysis within the Cosmoglobe Data Release 2 framework. The parametric form of the ZL model is inspired by the original DIRBE model by Kelsall et al. (K98), but the specific best-fit parameter values are re-derived using the combination of DIRBE Calibrated Individual Observations, Planck HFI sky maps, and WISE and Gaia compact object catalogs. Furthermore, the ZL parameters are fitted jointly with astrophysical parameters, such as thermal dust and starlight emission, and the new model takes into account excess radiation that appears stationary in solar-centric coordinates as reported in a companion paper. The relative differences between the predicted signals from K98 and our new model are $\lesssim 3\%$ in the 12 and 25 $μ$m channels over the full sky. The zero-levels of the cleaned DR2 maps are lower than those of the K98 ZL Subtracted Mission Average maps by $\sim 30$ kJy/sr at 1.25--3.5 $μ$m, which is larger than the entire predicted contribution from high-redshift galaxies to the Cosmic Infrared Background at the same wavelengths. At high Galactic latitudes, the total RMS of each DR2 map is lower than the corresponding DIRBE ZSMA map of $\sim$ 80 \% at wavelengths 4.9--25 $μ\mathrm{m}$. Still, obvious ZL residuals can be seen in several of the DR2 maps, and further work is required to mitigate these. Joint analysis with high-resolution full-sky surveys such as AKARI, IRAS, Planck HFI, and SPHEREx will be essential both to break key degeneracies in the current model and to determine whether the reported solar-centric excess radiation has a ZL or instrumental origin. Thus, while the results presented in this paper do redefine the state-of-the-art for DIRBE modelling, it also only represents the first among many steps toward a future optimal Bayesian ZL model. (abridged)