Bringing ExecuTorch To the Next Frontiers of Edge AI
summary
ExecuTorch 1.0 GA (Oct 2025) powers an ahead-of-time PyTorch export -> C++ runtime that runs on bare-metal ~50 KB runtimes, mobiles, M1 laptops and CUDA/Metal/Qualcomm backends; a live demo transcribes Vox Trail in real time fully locally on an M1. Microcontroller ARM results beat TFLite Micro on MobileNetV2 for latency/RAM/load time. Desktop work shows Qwen 3.5 running with 4-bit + turbo quant and custom Triton kernels at ~77 tokens/sec while aiming for CPU offload and Mixture-of-Experts support.
AI-written summary · about
chapters (11)
- 0:00ExecuTorch and edge AI
- 2:25ExecuTorch 1.0 capabilities
- 3:27Microcontroller support
- 5:11Voice in AI agents
- 8:22Real-time voice demo
- 10:53Inference on desktops
- 15:50Optimizing for GPUs
- 17:34Voice and desktop recap
- 18:58ExecuTorch open source
- 20:24PyTorch Foundation core
- 21:50Call to action
related talks
22mIntroducing ExecuTorch from PyTorch Edge: On-Device AI...
25mExecuTorch 1.0: General Availability Status for Mobile and Embedded...
33mExecutorch: Enabling and Accelerating PyTorch Models on Edge Devices
20mPyTorch Edge: Developer Journey for Deploying AI Models Onto Edge Devices
56mHow to: Bringing up ExecuTorch on a RISC-V Microcontroller
56mHow to: Bringing up ExecuTorch on a RISC-V Microcontroller | Corrected Audio
08mUnlocking AI at Scale for PyTorch Developers
16mLightning Talk: Efficient Inference at the Edge: Performance You Need at the Lowest...