PyTorch models now run directly on Cortex-M microcontrollers, under 90 KiB RAM

PyTorch · x · 2026-10-24

PyTorch and Arm jointly released the ExecuTorch Cortex-M backend, letting PyTorch models deploy directly on Arm Cortex-M microcontrollers without converting to another framework, accelerated via Arm CMSIS-NN. A single workflow covers all capability tiers (Scalar, DSP, MVE/Helium).

Key numbers:

On an Alif Ensemble E8 (single Cortex-M55 core), four of five MLPerf Tiny models run in ≤21.5 ms with under 90 KiB of RAM
The core runtime takes <20 KiB of flash; operator-selective builds make total code size scale with the model, not the framework
For heterogeneous SoCs it serves as the CPU execution path alongside Ethos-U NPU delegation

Previously, MCU deployment (<1 MB SRAM, milliwatt power budgets) required cross-framework conversion, separate toolchains, and debugging numerical mismatches — all now eliminated.

Original post →

PyTorch models now run directly on Cortex-M microcontrollers, under 90 KiB RAM

More from Infra