ExecuTorch is PyTorch’s open-source solution for running AI models on-device, across mobile phones, embedded systems and edge hardware. Maintained by the PyTorch team at Meta, the GitHub repository has more than 5,100 stars and is under very active development, with updates continuing through 2026. It is the deployment counterpart to PyTorch training: you export a trained model and run it efficiently on the device itself, with no server round-trip.
ExecuTorch takes a PyTorch model, exports it through the torch.export pipeline, lowers it for a specific hardware backend, and produces a compact .pte program file that a lightweight C++ runtime can execute. It is the successor path to the older TorchScript mobile stack, designed from the start for the generative AI era, where models are larger and hardware accelerators far more diverse. The runtime is designed for constrained environments, with a small binary footprint and portable kernels, while backend delegates let chips from Apple, Arm, Qualcomm, MediaTek and others accelerate inference with their own NPUs, GPUs and DSPs.
Free and open source from the PyTorch project. There is no paid tier or hosted service; you build and ship the runtime inside your own applications.
Mobile and embedded developers who train in PyTorch and need private, low-latency, offline inference on end-user devices, from smartphone AI features to industrial edge hardware.