The MIT vision library treating training as a feature, not an upsell

What it does LibreYOLO is a Python computer vision library that exposes detection, segmentation, pose estimation, depth prediction, OCR, and roughly a dozen other tasks through a single class. Swap the checkpoint—LibreYOLO9t.pt for detection, LibreHRNetw32-pose.pt for keypoints, LibreMiDaSs-depth.pt for depth—and the calling code stays identical. It reads standard YOLO-format datasets, so existing training pipelines port over without reformatting your data.

The interesting bit The project acts as a curated, MIT-licensed integration layer over a sprawling zoo of models—YOLOv9, RF-DETR, SAM 3, Depth Anything 3, Florence-2, and many others—unified by one API and one CLI. The README explicitly notes that training and export are “included rather than sold separately,” which frames it as a permissive alternative to libraries that gate those capabilities behind commercial licenses.

Key highlights

• One API covers seventeen tasks, from oriented bounding boxes to gaze estimation and background removal.
• Supports a wide model zoo including YOLOv9, RT-DETR variants, SAM/EdgeTAM, CLIP/SigLIP2, and several VLMs.
• Exports to twelve formats: ONNX, TensorRT, OpenVINO, CoreML, TFLite, NCNN, MNN, Paddle, ExecuTorch, and others.
• Training supports multi-GPU, LoRA, distillation, and logging to TensorBoard, Weights & Biases, MLflow, and more.
• Pre-trained weights inherit their original licenses, so not every checkpoint is commercially permissive despite the MIT code.

Caveats

• Export support varies significantly by model family and task; the README directs users to a compatibility matrix rather than promising universal conversion.
• Pre-trained weights carry their original licenses, meaning the MIT code license does not guarantee you can use every checkpoint commercially.
• Several optional extras (executorch, coreai, neptune) are deliberately excluded from the all install target because they pin torch or protobuf aggressively.

Verdict Worth a look if you want a permissive, MIT-licensed alternative to Ultralytics with broad task coverage and standard YOLO dataset compatibility. Skip it if you need a single, battle-hardened model family with guaranteed export parity across every format.