How Meta accelerates on-device ML on its apps with ExecuTorch
Original post Accelerating on-device ML on Meta’s apps with ExecuTorch
Article Summary
Meta just shared real numbers on their shift to on-device ML across billions of users. The PyTorch Edge Team reveals how ExecuTorch is powering Instagram, WhatsApp, Messenger, and Facebook with faster, more private AI.
Meta has rolled out ExecuTorch, their open-source PyTorch inference framework, across their family of apps over the past year. Built with industry partners like Arm, Apple, and Qualcomm, it's designed to run ML models directly on mobile devices for better latency, privacy, and offline functionality.
Key Takeaways
- Instagram Cutouts saw DAU increases after migrating to ExecuTorch with SqueezeSAM
- WhatsApp reduced model load time and inference time while cutting ANR metrics
- Messenger enabled end-to-end encryption by moving server-side models on-device
- Facebook's SceneX model shows performance improvements across all device tiers
ExecuTorch delivered measurable performance gains and enabled new privacy-preserving features across Meta's apps serving billions of users.
Continue reading
This page summarizes “Accelerating on-device ML on Meta’s apps with ExecuTorch” by Meta, published Jul 28, 2025. All credit for the original work goes to the authors.