ExecuTorch

What is ExecuTorch?

ExecuTorch is the on-device AI framework from the PyTorch ecosystem. It lets you export a PyTorch model directly, without converting to a separate intermediate format, then compile it ahead of time for the hardware you target and run it with a small C++ runtime. The core runtime has a base footprint of about 50KB, and the same exported .pte file can run on microcontrollers up to smartphones. You choose from more than 12 hardware backends, including XNNPACK for Arm CPUs, Core ML for Apple devices, the Qualcomm AI Engine, Vulkan, MediaTek and Samsung, by changing a single partitioner line. Native APIs are available for C++, Swift, Kotlin, Objective-C and WebAssembly. Because inference stays on the device, text, vision and audio features can work offline, keep data private and avoid cloud compute bills. The framework is fully open source, and an experimental in-browser playground runs models on WebGPU.

Use cases

• Deploy PyTorch models to mobile apps
• Run language models on-device
• Add offline speech-to-text to an app
• Run vision models on edge hardware
• Quantize models for small devices
• Target Apple, Qualcomm or Arm hardware
• Prototype models in the browser with WebGPU

Who it's for

• ML engineers
• Mobile developers
• Embedded developers
• AI researchers

Why use ExecuTorch

• Models stay in the PyTorch toolchain, so there is no need to rewrite them in C++ or convert to vendor formats.
• On-device inference keeps personal data on the device and works without a network connection.
• No cloud compute bills or API rate limits, so costs do not grow with each added user.
• Switching hardware backends is a one-line change, which eases support for several device families.
• The page states it is used in production for billions of users at Meta.

ExecuTorch features

• Direct export from PyTorch with core ATen operators preserved
• Ahead-of-time compilation tuned to the target device
• 12+ hardware backends such as XNNPACK, Core ML and Qualcomm
• Portable C++ runtime from microcontrollers to smartphones
• 50KB base runtime footprint
• Native APIs for C++, Swift, Kotlin, Objective-C and WebAssembly
• Quantization through torchao
• Modular, composable optimization steps
• Browser playground for chat, speech-to-text, camera vision and segmentation on WebGPU

ExecuTorch integrations

• PyTorch
• torchao
• XNNPACK
• Arm Kleidi
• Core ML
• Qualcomm Hexagon
• Vulkan
• MediaTek
• Samsung
• WebGPU

How to get started with ExecuTorch

1. Export your PyTorch model with torch.export to produce a graph equivalent to the original.
2. Lower it for your target hardware with to_edge_transform_and_lower and a backend partitioner such as XNNPACK, Core ML or Qualcomm.
3. Save the result as a .pte file using to_executorch.
4. Load the .pte file in your app through the C++, Swift, Kotlin, Objective-C or WebAssembly API and call forward.
5. Try the browser playground first if you want to see models run on WebGPU without installing anything.

ExecuTorch limitations

• The browser playground is labeled experimental
• Developers must pick a backend and export pipeline for each target device
• Edge power, thermal and memory limits still constrain which models fit

ExecuTorch pricing

ExecuTorch is open source, so you can use and self-host it for free.

Prices can change and may vary by country or exclude tax. Check the official pricing page.

ExecuTorch alternatives

The main ExecuTorch alternatives listed on Debuly are llama.cpp, LlamaEdge, Nota AI, TensorFlow, bitnet.cpp and DeepEdge. The table compares pricing and platforms; order is A–Z.

1. 

   bitnet.cpp

   Microsoft's official open-source inference framework for 1-bit LLMs, with optimized kernels for running BitNet b1.58 models on CPUs and GPUs.

   Developer platforms
   Open sourceVisit bitnet.cpp (official site)
2. 

   DeepEdge

   Low-code deep learning platform to label, train, optimize and deploy computer vision models on edge hardware.

   Developer platforms
   Free trialVisit DeepEdge (official site)
3. 

   llama.cpp

   Open-source engine and desktop app for running LLMs locally, with a built-in OpenAI-compatible API server and web chat UI.

   Developer platformsCodingChat assistants
   Open sourceVisit llama.cpp (official site)
4. 

   LlamaEdge

   Small Rust and WebAssembly runtime that serves local LLMs through an OpenAI-compatible API on CPUs, GPUs and edge devices.

   Developer platformsCodingAutomation and agents
   Open sourceVisit LlamaEdge (official site)
5. 

   Nota AI

   Hardware-aware AI model optimization platform, NetsPresso, plus edge AI vision products for traffic, in-cabin monitoring and video analysis.

   Developer platformsAutomation and agents
   Visit Nota AI (official site)
6. 

   TensorFlow

   Open-source machine learning framework with Keras, TensorFlow.js, LiteRT and TFX for building and deploying models.

   Developer platformsCoding
   Open sourceVisit TensorFlow (official site)

ExecuTorch FAQ

Sources & verification

Facts on this page come from the official sources below, each with the date we read it. Written by Debuly; the vendor did not review this page.

Image credits

• Logo: from raw.githubusercontent.com · fetched 23 Oct 2026 · Vendor trademark, used only to identify the product; removed on request
• Cover: none collected yet, so a generated tile is shown.

Spot an error? Report an issue with this listing. We also remove a logo or image on the owner's request.