ExecuTorch vs PyTorch Mobile: The Next-Gen On-Device Training Runtime

[PyTorch Mobile] excels at providing a stable, mature runtime for deploying models on Android and iOS because it leverages a well-established operator set and a straightforward model conversion path from standard PyTorch. For example, teams with existing torch.jit.trace workflows can deploy to mobile with minimal refactoring, achieving reliable inference on millions of devices. However, its architecture was not designed for the extreme size and performance constraints of modern microcontrollers or the full delegate composability required for heterogeneous federated learning fleets.

[ExecuTorch] takes a fundamentally different approach by being a ground-up redesign focused on portability, minimal binary size, and first-class support for custom hardware accelerators via a unified delegate API. This results in a runtime that can target everything from high-end smartphones to low-power Cortex-M microcontrollers, enabling a single toolchain for diverse on-device training tasks. The trade-off is that it is a newer, evolving platform, requiring teams to migrate from the legacy torch.jit ecosystem to a new export-based workflow.

The key trade-off: If your priority is maintaining a stable, battle-tested mobile inference pipeline with minimal migration cost, choose [PyTorch Mobile]. If you prioritize a future-proof, ultra-portable runtime that can unify on-device training across smartphones, wearables, and microcontrollers for a truly heterogeneous federated learning system, choose [ExecuTorch].

ExecuTorch excels at deploying highly optimized, specialized models on heterogeneous edge hardware because of its delegate-centric architecture. By separating the runtime core from hardware-specific backends, it enables aggressive binary size reduction (often under 2MB for a minimal build) and unlocks custom accelerators like DSPs and NPUs that PyTorch Mobile cannot easily target. For example, early adopters report a 30-50% reduction in binary footprint compared to equivalent PyTorch Mobile builds, a critical factor for IoT and micro-controller deployments.

PyTorch Mobile takes a different approach by providing a mature, monolithic runtime that closely mirrors the server-side PyTorch experience. This results in a much lower migration friction for existing pipelines and broader out-of-the-box operator coverage without requiring custom delegate authoring. The trade-off is a larger minimum binary size and a less modular architecture, making it harder to strip unused operators for ultra-constrained environments.

The key trade-off: If your priority is maximum portability across novel silicon and minimal binary size for a new federated learning fleet, choose ExecuTorch. If you prioritize immediate compatibility with an existing PyTorch Mobile codebase and a stable, battle-tested operator set, choose PyTorch Mobile for now, but plan a phased migration as ExecuTorch's ecosystem matures.