Thorium‐234 (234Th), an insoluble radioisotope scavenged by marine particles, can be used as a proxy of the biological carbon pump. Thorium‐234 observations can constrain biogeochemical models, but a necessary first step is to estimate the poorly known partition coefficients between particulate and dissolved phases. In this study, the 234Th partition coefficients for five particle types, differing in size and chemical composition, are estimated by fitting a global 3‐D 234Th model based on the coupled ocean general circulation‐biogeochemistry model NEMO‐PISCES (at a resolution of 2°) to a global 234Th data set (including GEOTRACES data). Surface partition coefficients are estimated between 0.79 and 16.7×106. Biogenic silica has the smallest partition coefficients. Small particulate organic carbon and lithogenic dust have the largest. Thorium‐234 observations at depth cannot be recovered without allowing partition coefficients to increase by one order of magnitude from surface to 1,000 m deep. In our time‐dependent global 3‐D model, the biases introduced by three common assumptions made in biological carbon pump studies can be quantified. First, using the C:234Th ratio of large particles alone leads to an overestimation of carbon export at the base of the euphotic layer, by up to a factor 2. Furthermore, assuming steady state and neglecting transport by advection and diffusion can bias fluxes by as much as 50%, especially at high latitudes and in upwellings, with a sign and intensity depending on the season.