ExecuTorch

On-device inference for PyTorch models is ExecuTorch's whole job: it exports a model straight from PyTorch, compiles it ahead of time, and runs it on phones, wearables, and embedded boards through a runtime with a base footprint of about 50 KB. There is no intermediate format conversion; the same exported graph flows to one of a dozen hardware backends including XNNPACK for CPU, Core ML, Vulkan, Qualcomm QNN, MediaTek, Samsung Exynos, ARM Ethos-U, OpenVINO, NXP, and Cadence DSP, with experimental CUDA and Metal paths, quantization handled through torchao, and memory planning plus selective builds keeping binaries small. Installation is pip install executorch for the Python authoring side, with C++ and mobile SDKs for deployment. Meta runs it in production across Instagram, WhatsApp, Quest 3, and Ray-Ban Meta smart glasses, making it one of the most widely deployed edge inference runtimes anywhere. The code is BSD-3-Clause licensed under the PyTorch organization, with documentation at docs.pytorch.org/executorch.