---

name: export-and-deployment description: "Export, benchmark, and route Ultralytics YOLO models to deployment formats including ONNX, OpenVINO, TensorRT, CoreML, TFLite, Triton, and edge runtimes." disable-model-invocation: true metadata: disco-role: operating license: AGPL 3.0

Export And Deployment

Use this sub-skill when the user needs model.export(), yolo export, yolo benchmark, backend dependency choices, or deployment-format routing for Ultralytics YOLO models. It covers export/deploy decisions only; route data YAML setup to data-and-configuration, training and validation to training-and-validation, prediction result handling to inference-and-results, model-family selection to model-families-and-tasks, and repo/test maintenance to repo-development.

Fast Routing

• Portable CPU/server deployment: prefer format=onnx; it is handled by the export-base extra and can be used through ONNX Runtime, OpenCV DNN (dnn=True), or Triton.
• Intel CPU/GPU/NPU deployment: prefer format=openvino; use device=intel:cpu, device=intel:gpu, or device=intel:npu when running exported models.
• NVIDIA GPU deployment: prefer format=engine for TensorRT; export and inference require compatible CUDA/NVIDIA runtime, and INT8 calibration should happen on the target GPU class.
• Apple apps: prefer format=coreml; export on macOS or x86 Linux, but Ultralytics CoreML predict/val execution is macOS-only.
• Mobile/embedded TensorFlow: prefer format=tflite or format=edgetpu; these use TensorFlow export tooling and often require pinned dependency environments.
• Benchmark comparison: use yolo benchmark model=... data=... imgsz=... format=... for one target, or omit format to attempt all supported formats that work in the current environment.

Command Shapes

Ultralytics CLI commands use yolo TASK MODE arg=value; TASK is optional for export/benchmark when the model implies it.

yolo export model=yolo26n.pt format=onnx imgsz=640 dynamic=True
yolo export model=best.pt format=engine device=0 half=True workspace=4
yolo benchmark model=yolo26n.pt data=coco8.yaml imgsz=640 format=onnx
yolo predict model=yolo26n.onnx source=image.jpg

Python API equivalents:

from ultralytics import YOLO

model = YOLO("yolo26n.pt")
path = model.export(format="onnx", imgsz=640, dynamic=True)
metrics = YOLO(path).val(data="coco8.yaml")

Before suggesting a command that might download weights, export large artifacts, pull containers, train, benchmark all formats, or touch media, call out the side effect and offer a dry-run plan first. For a safe local plan, use scripts/plan_export.py.

Important Export Arguments

• format: export target such as onnx, openvino, engine, coreml, saved_model, pb, tflite, edgetpu, tfjs, mnn, ncnn, imx, rknn, executorch, axelera, deepx, or qnn.
• imgsz: export input size; benchmark mode requires square image size.
• device: CPU/GPU/backend selector; TensorRT defaults toward GPU export, OpenVINO inference accepts Intel device strings, and CoreML export may use CPU/MPS/GPU depending on platform.
• half and int8: mutually exclusive in exporter behavior; int8=True needs representative calibration data for reliable accuracy.
• data and fraction: calibration dataset and subset fraction for INT8 formats such as ONNX, OpenVINO, TensorRT, TFLite, Edge TPU, IMX, RKNN, Axelera, DeepX, and QNN.
• dynamic, batch, nms, opset, simplify, workspace, keras, and optimize: format-specific knobs; verify support before adding them.

See references/workflows.md for recipes and references/api-reference.md for the format matrix. See references/troubleshooting.md when export fails before recommending broad dependency reinstallations.

Safe Planning Script

Use the bundled script to validate a requested target without running export:

python sub-skills/export-and-deployment/scripts/plan_export.py --model yolo26n.pt --format onnx --dynamic --int8 --data coco8.yaml
python sub-skills/export-and-deployment/scripts/plan_export.py --model best.pt --format engine --device cpu

The script emits JSON with the planned yolo export command, supported arguments, dependency extra notes, and warnings such as CPU-only TensorRT or ONNX export-base requirements.