Libreyolo

LibreYOLO is an MIT licensed Python library for training, running and exporting object detection models. It presents one API, called with a single LibreYOLO() call, that sits in front of YOLOX, YOLOv9 and RF DETR, with architecture, size and class count detected automatically from the weights. The project positions itself as an alternative to Ultralytics, with no AGPL code anywhere in its dependency chain, so it can be used in closed source products, SaaS and embedded systems. The library accepts paths, URLs, PIL images, NumPy arrays, OpenCV frames, tensors and bytes as input, supports tiled inference for large images, and downloads weights from HuggingFace. Users can fine tune on custom YOLO or COCO datasets with built in augmentation, mixed precision, early stopping and checkpoint resume, and validate with COCO standard metrics including mAP50, mAP50 to 95, precision, recall, per class metrics and a confusion matrix. Models export to ONNX with embedded metadata. Beyond detection, the site describes one API covering 20 tasks, including segmentation, keypoints, classification, depth, text recognition, embeddings, matting and 3D detection, with segmentation alone listing 14 models such as SAM, SAM 2 and SAM 3. LibreYOLO is free and open source. The site publishes no paid plans, and support is through GitHub, issues and a community subreddit.