ExecuTorch
PyTorch's open-source stack for running AI models locally on phones, wearables, laptops, browsers, embedded systems and microcontrollers.
ExecuTorch starts from a PyTorch model, captures it with torch.export, optimizes it for target hardware, and runs it through C++, Python, Swift/Objective-C, Kotlin/Java or JavaScript APIs. It already powers on-device features in Instagram, WhatsApp, Facebook, Messenger, Meta Quest and Ray-Ban Meta devices.
It covers local LLM and agent building blocks such as quantized text models, long context, tool calling and speculative decoding; voice tasks such as streaming and offline speech recognition, speech synthesis, voice-activity detection and speaker diarization; multimodal text, image and audio runners and mobile vision-language models; computer vision tasks such as classification, detection and segmentation; and embedded AI on Cortex-M CPUs, Ethos-U and NXP NPUs, Cadence DSPs, Zephyr, Arduino and Raspberry Pi Pico.
It is written in Python; a model does not need to appear in its listed examples to run, since deployment depends on torch.export capture plus operator, runtime and backend coverage that a developer validates for the target hardware.