A modified version of the Biome4 vegetation model for simulation of the mean delta(13)C of plant communities is presented, and used to reconstruct palaeoprecipitation. We treat all fractionations by C3 and C4 plants in all coexistent Plant Functional Types, weighted by their respective net primary production. We constrain the range of variation in the intracellular versus atmospheric CO2 concentration by fixing a lower limit. Finally, we replace some constant parameters by functions of external forcing to account for their responses to environmental variation. The new version of Biome4 was applied as an inverse model and tested on three modern data sets. The fit between observations and simulations is very close to the 1:1 relationship, with respective slopes of 0.90 +/- 0.02 (r(2) = 0.98, n=29) for delta(13)C and 0.97 +/- 0.06 (r(2) = 0.90, n = 29) for precipitation. Inverse modelling was applied using the Metropolis-Hastings algorithm to the NuBloch loess sequence. Over the last glaciation, simulated palaeoprecipitation varies between 240 mm year(-1) and 400 min year(-1). This study clearly demonstrates atmospheric teleconnections with the Greenland ice-sheet extension, by matching Dansgaard-Oeschger events with precipitation increase of ca. 100-200 min year(-1).