Advancements in mode-division multiplexing (MDM) techniques, aimed at surpassing the Shannon limit and augmenting transmission capacity, have garnered significant attention in optical fiber communication, propelling the demand for high-quality multiplexers and demultiplexers. However, the criteria for ideal-mode multiplexers/demultiplexers, such as performance, scalability, compatibility, and ultra-compactness, have only partially been achieved using conventional bulky devices (e.g., waveguides, gratings, and free space optics)—an issue that will substantially restrict the application of MDM techniques. Here, we present a neuro-meta-router (NMR) optimized through deep learning that achieves spatial multi-mode division and supports multi-channel communication, potentially offering scalability, compatibility, and ultra-compactness. An MDM communication system based on an NMR is theoretically designed and experimentally demonstrated to enable simultaneous and independent multi-dataset transmission, showcasing a capacity of up to 100 gigabits per second (Gbps) and a symbol error rate down to the order of 10⁻⁴, all achieved without any compensation technologies or correlation devices. Our work presents a paradigm that merges metasurfaces, fiber communications, and deep learning, with potential applications in intelligent metasurface-aided optical interconnection, as well as all-optical pattern recognition and classification.