PurposeThe aim of this work was to develop and benchmark a magnetic resonance (MR)‐guided linear accelerator head model using the GEANT4 Monte Carlo (MC) code. The validated model was compared to the treatment planning system (TPS) and was also used to quantify the electron return effect (ERE) at a lung–water interface.MethodsThe average energy, including the spread in the energy distribution, and the radial intensity distribution of the incident electron beam were iteratively optimized in order to match the simulated beam profiles and percent depth dose (PDD) data to measured data. The GEANT4 MC model was then compared to the TPS model using several photon beam tests including oblique beams, an off‐axis aperture, and heterogeneous phantoms. The benchmarked MC model was utilized to compute output factors (OFs) with the 0.35 T magnetic field turned on and off. The ERE was quantified at a lung–water interface by simulating PDD curves with and without the magnetic field for 6.6 × 6.6 cm2 and 2.5 × 2.5 cm2 field sizes. A 2%/2 mm gamma criterion was used to compare the MC model with the TPS data throughout this study.ResultsThe final incident electron beam parameters were 6.0 MeV average energy with a 1.5 MeV full width at half maximum (FWHM) Gaussian energy spread and a 1.0 mm FWHM Gaussian radial intensity distribution. The MC‐simulated OFs were found to be in agreement with the TPS‐calculated and measured OFs, and no statistical difference was observed between the 0.35 T and 0.0 T OFs. Good agreement was observed between the TPS‐calculated and MC‐simulated data for the photon beam tests with gamma pass rates ranging from 96% to 100%. An increase of 4.3% in the ERE was observed for the 6.6 × 6.6 cm2 field size relative to the 2.5 × 2.5 cm2 field size. The ratio of the 0.35 T PDD to the 0.0 T PDD was found to be up to 1.098 near lung–water interfaces for the 6.6 × 6.6 cm2 field size using the MC model.ConclusionsA vendor‐independent Monte Carlo model has been developed and benchmarked for a 0.35 T/6 MV MR‐linac. Good agreement was obtained between the GEANT4 and TPS models except near heterogeneity interfaces.