The predictive capability of two comprehensive combustion codes, PCGC-3 and FLUENT, to simulate local flame structure and combustion characteristics in an industrial gas-fired, flat-glass furnace is investigated. Model predictions are compared with experimental data from the furnace for profiles of velocity, species concentrations, temperatures, and wall-incident radiative heat flux. Predictions from both codes show agreement with the measured mean velocity profiles and incident radiant flux on the crown. However, significant differences between the code predictions and measurements are observed for the flame-zone temperatures and species concentrations. The observed discrepancies may be explained by (i) uncertainties in the distributions of mean velocity and turbulence in the portneck, (ii) uncertainties in the port-by-port stoichiometry, (iii) different grid-based approximations to the furnace geometry made in the two codes, (iv) the assumption of infinitely fast chemistry made in the chemical reaction model of both codes, and (v) simplifying assumptions made in the simulations regarding the complex coupling between the combustion space, batch blanket, and melt tank. The study illustrates the critical need for accurate boundary conditions (inlet air and fuel flow distributions, boundary surface temperatures, etc.) and the importance of representative furnace geometry in simulating these complex industrial combustion systems.