There has been considerable success applying deep learning to cardiac MRI segmentation. However, segmentation masks have several shortcomings. In particular, they are discrete (voxel-based) representations of continuous anatomy, and are hence not suitable for use in biomechanical simulation. Mesh representations can potentially overcome both of these drawbacks. We demonstrate an approach to predicting left-ventricular (LV) meshes from short-axis MR images. The proposed approach: (i) works robustly on data with differing slice numbers, slice thicknesses, and left-ventricular coverage, (ii) has no test-time optimisation loop, but rather directly predicts the mesh from a mask, and, (iii) generalises to real data with pathology.