👋 Hello @ambitious-octopus, thank you for submitting an ultralytics/ultralytics 🚀 PR! This is an automated response — an Ultralytics engineer will review and assist you shortly. To ensure a seamless integration of your work, please review the following checklist:

• ✅ Define a Purpose: Clearly explain the purpose of your fix or feature in your PR description, and link to any relevant issues. Ensure your commit messages are clear, concise, and adhere to the project's conventions.
• ✅ Synchronize with Source: Confirm your PR is synchronized with the ultralytics/ultralytics main branch. If it's behind, update it by clicking the 'Update branch' button or by running git pull and git merge main locally.
• ✅ Ensure CI Checks Pass: Verify all Ultralytics Continuous Integration (CI) checks are passing. If any checks fail, please address the issues.
• ✅ Update Documentation: Update the relevant documentation for any new or modified features.
• ✅ Add Tests: If applicable, include or update tests to cover your changes, and confirm that all tests are passing.
• ✅ Sign the CLA: Please ensure you have signed our Contributor License Agreement if this is your first Ultralytics PR by writing "I have read the CLA Document and I sign the CLA" in a new message.
• ✅ Minimize Changes: Limit your changes to the minimum necessary for your bug fix or feature addition. "It is not daily increase but daily decrease, hack away the unessential. The closer to the source, the less wastage there is." — Bruce Lee

Additional notes for this PR:

• Please confirm the new executorch export target is consistently listed in export_formats() and correctly wired in the exporter flow end-to-end.
• Double-check naming consistency (format string, suffix, and user-facing messages) and that the new method is covered by unit tests, including edge cases like dynamic shapes or invalid inputs.
• Ensure dependency checks (executorch>=0.7.0) are properly guarded and error messages are clear if the requirement isn’t met.
• If there are known limitations (e.g., fixed input shape), kindly document them in code comments and in the user docs update.
• Consider adding a brief usage snippet to the export docs to align with existing exporter UX, matching the style of other formats.

For more guidance, please refer to our Contributing Guide. Don’t hesitate to leave a comment if you have any questions. Thank you for contributing to Ultralytics! ⚡📱

The reason we added setuptools because until 0.7.0 there is a usage of "pkg_resources" and it is about to be removed in setuptools and there is a deprecation warning in code as well. But good news is in recent PR it is already fixed. : pytorch/executorch#13513 (it should be available in next release)

When new package release we can remove setuptools as well.

cc @ambitious-octopus

I also found one left over pkg_resource import in samsung backend : pytorch/executorch#14654
Hopefully it will get merged as well.

Codecov Report

❌ Patch coverage is 64.10256% with 14 lines in your changes missing coverage. Please review.

📢 Thoughts on this report? Let us know!

from ultralytics import YOLO

model = YOLO("yolo11n.pt")

executorch_path = model.export(format="executorch")
model = YOLO(executorch_path)

model("https://ultralytics.com/images/bus.jpg",task='detect',imgsz=640, conf=0.25, save=True)

@ambitious-octopus , I made some changes on better export and also autobackend part. I also had so make some changes on suffix part to avoid ".pt" vs ".pte" cases. Plus I had to add "yaml" file for metadata for proper "class names" in inference

Great new export addition guys! @ambitious-octopus @onuralpszr

Here is performance testing on Raspberry Pi 5 16GB:

Results on COCO128 with YOLO11n:

• Baseline PyTorch: 373.1ms inference
• ExecuTorch + XNNPACK: 184.2ms inference
• Speedup: ~2x faster

Note: It takes about 20 minutes to compile ExecuTorch from source on a Raspberry Pi 5 16GB.

I have also investigated about aarch64 Linux wheels for Executorch on Executorch GitHub and have come across the following PR which has added the wheel building support on a nightly release cadence but not uploading to PyPi.
pytorch/executorch#13946

So I have gone ahead and tried the latest nighly release from below with Python3.11 support which comes with Raspberry Pi OS Bookworm (Debian 12):
https://github.com/pytorch/executorch/actions/runs/18128543623

Installation of the above wheel resulted in a missing module:

ModuleNotFoundError: No module named 'torchao'

But I was able to resolve it by pip install torchao (torchao has PyPi aarch64 wheel) and then able to successfully export a model with ExecuTorch.

So I think we have 2 options here:

1. Host a compiled Executorch aarch64 Linux wheel on our GitHub assets.
2. Discuss with ExecuTorch guys and maybe Raspberry Pi guys as well to have the official aarch64 Linux wheel support on PyPi.

@lakshanthad personally, I wanna see "pypi" push first instead of assets, secondly we need to think about cpp examples because better inference speed and most of the example in their docs also revolving around cpp as well.

I understand @onuralpszr. However PyPi push has been discussed in multiple issues and seems the wheels never made it there.

• Request for support of ExecuTorch pip package on linux aarch64 pytorch/executorch#10651
• Publish Linux ARM64 Wheels to PyPI pytorch/executorch#7331
• Publish pypi wheel for linux aarch64 pytorch/executorch#13044

Yeah definitely we can experiment around CPP as well!

⚡ Actions Trigger

_{Made with ❤️ by Ultralytics Actions}

GitHub Actions below triggered via workflow dispatch on this PR branch executorch-export at 2025-10-21 20:44:17 UTC with @ultralytics/run-ci command:

• ✅ CI: ci.yml run 13463
• ✅ Publish Docker Images: docker.yml run 1320

@ambitious-octopus @lakshanthad must torch 2.9 because of TorchAO 0.14 needs 2.9 (pytorch/ao#2919)

@glenn-jocher had to make some extra adjustment because of torch 2.9.0 but it is ready for initial merge, tests are should passed, slow CI last minute changes because of docker side was using torch 2.8 make me re-check dep and I make sure executorch using correct packages and make sure match with supported with TorchAO as well.

cc @ambitious-octopus

Thanks @onuralpszr, looking through this now

🎉 Merged! Huge thanks to @ambitious-octopus for leading PR #22244, with stellar contributions from @larryliu0820, @lakshanthad, @Y-T-G, @onuralpszr, @mergennachin, and @glenn-jocher. This brings first-class ExecuTorch support to YOLO11—unlocking seamless mobile and edge deployment with a unified CLI/Python experience, solid docs, and reliable tests.

“Innovation is the ability to see change as an opportunity — not a threat.” — Steve Jobs
Your work turns opportunity into reality: export once, run anywhere, and accelerate CPU inference on-device. 🚀

Quickstart:

• CLI: yolo export model=yolo11n.pt format=executorch && yolo predict model=yolo11n_executorch_model/model.pte
• Python:

from ultralytics import YOLO

model = YOLO("yolo11n.pt")
model.export(format="executorch")
YOLO("yolo11n_executorch_model/model.pte")(source="test.jpg")

Appreciate the craftsmanship, polish, and care—this meaningfully grows the Ultralytics ecosystem and empowers developers from Ultralytics HUB to embedded devices. Thank you all! 🙌