ExecuTorch
PyTorch-native framework for running AI models on phones, wearables, and microcontrollers.
About
ExecuTorch is PyTorch's open-source stack for on-device AI inference, developed by the PyTorch team at Meta. It takes a PyTorch model via torch.export, optimizes it for the target hardware, and runs it through C++, Python, Swift, Kotlin/Java, or JavaScript APIs with no manual rewrites or format conversions. Version 1.0 shipped in October 2024.
It already powers on-device AI across Instagram, WhatsApp, Facebook, and Messenger, serving billions of users, plus Meta Quest and Ray-Ban Meta devices. Hardware delegates cover CPU, GPU, and NPU through backends like XNNPACK and Core ML, and the same APIs handle LLMs, vision-language models, image segmentation, detection, and audio.
Honest limitations: on-device deployment still means real constraints on model size and latency budgets, and squeezing a model onto a phone often requires quantization and delegate-specific tuning. The toolchain is powerful but has a learning curve around export, delegation, and debugging compared to cloud inference.
What it can do
- PyTorch-native on-device inference from phones to microcontrollers
- Hardware delegates for CPU, GPU, and NPU (XNNPACK, Core ML, and more)
- LLM, vision, speech, and multimodal model support with one export flow
Setup
pip install executorch, export a PyTorch model with torch.export, then build and run via the Python or native APIs.
Prerequisites
A PyTorch model and familiarity with model export; target device SDKs (iOS/Android) for mobile deployment.
Use cases
Works with
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 81; Trending score 88/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.