To solve on-chip beam combining losses in optical neural networks, the authors introduce a multimode photonic convolver via lossless mode division fan-in. With mode and wavelength multiplexing, it achieves 6–7 bit precision and a computational density of 125 TOPS/mm2.