Every YOLO variant, but make it Paddle

What it does PaddleYOLO is a specialized offshoot of PaddleDetection that contains nothing except YOLO-series object detectors and their Paddle-native cousins. It hosts configs, weights, and training pipelines for a sprawling roster: legacy YOLOv3, the in-house PP-YOLOE+ line, third-party ports through YOLOv11, plus YOLOX, RT-DETR, RTMDet, and YOLOv6Lite. The repo covers the full model lifecycle—training, evaluation, inference, ONNX export, and TensorRT benchmarking—while remaining tightly coupled to the PaddlePaddle runtime.

The interesting bit Rather than letting YOLO code bloat the main PaddleDetection repository, the maintainers deliberately quarantined it here under a GPL 3.0 license. That means popular models such as YOLOv5 and YOLOv8 live exclusively in this annex and will never merge back upstream, making this the official YOLO embassy for the Paddle ecosystem.

Key highlights

• Exhaustive model roster: YOLOv3, PP-YOLO/PP-YOLOE/PP-YOLOE+, YOLOX, YOLOv5 through YOLOv11, YOLOv5u/v7u, YOLOv6Lite, RTMDet, and RT-DETR
• Swap-in backbones including ConvNeXt, Swin-Transformer, ViT, and FocalNet for PP-YOLOE+ variants
• Quantization-aware training via PaddleSlim and multi-hardware deployment through FastDeploy
• Export paths to ONNX and TensorRT with optional post-processing stripped for custom inference pipelines
• Dedicated Model Zoo with COCO-pretrained weights and VOC or custom-dataset recipes

Caveats

• The GPL 3.0 license is stricter than typical Apache-2.0 ML tooling, which may complicate commercial downstream use
• Several popular models are explicitly noted as never merging into PaddleDetection, so this repo is their only official Paddle home
• Windows users are directed to install a paddle develop build rather than a stable release, suggesting platform-specific rough edges

Verdict Worth bookmarking if you are already committed to the PaddlePaddle stack and need a one-stop shop for YOLO research or production. Everyone else already running PyTorch-native YOLOv8 should probably keep scrolling.