Meta AI Updates: October 17, 2026

1. ExecuTorch Adds a Cortex-M Backend for Running PyTorch Models on Microcontrollers

Meta and Arm. ExecuTorch now has a Cortex-M CPU backend that swaps standard PyTorch operators for Arm CMSIS-NN kernels through ahead-of-time compiler passes, covering Cortex-M0 through M85 and the Scalar, DSP, and MVE/Helium ISA tiers from one export, quantize, and lower workflow. On an Alif Ensemble E8, four of the five MLPerf Tiny models each ran in 21.5 ms or less using under 90 KiB of RAM on a single Cortex-M55 core, and the core runtime takes under 20 KiB of flash, with selective build linking only the operators a model uses. On SoCs with an Ethos-U NPU, the backend provides the CPU fallback path alongside NPU delegation. For teams training in PyTorch, this removes the need to convert models to another framework format for MCU deployment. Source