Deeptanshu-sankhwar/horizonml

HorizonML attempts to solve a complex engineering problem—distributed model parallelism on resource-constrained edge devices—but lacks the traction and development velocity to be considered a viable contender in the space. With only 10 stars and zero forks over 431 days, the project appears to be a stagnant academic or personal experiment rather than an active framework. Technically, the challenge of model parallelism (splitting a model across devices) on edge hardware is significantly hampered by network latency and bandwidth, issues that established projects like FedML, Flower (for federated learning), and PyTorch's own RPC/Distributed features address with much larger engineering teams and community support. The 'heterogeneous' aspect is a known research challenge, but without active maintenance or a breakthrough in communication efficiency, this project is easily displaced by more robust ecosystem players. Platform risk is high because as edge training becomes more viable, hardware-software vertical integrators like Apple (MLX/CoreML) and Google (TensorFlow Lite/Edge TPU) will likely provide native, optimized primitives that render third-party edge training wrappers obsolete.