YOLO-World

Tencent AI Lab's YOLO-World, a CVPR 2024 paper, brought open-vocabulary detection to real-time speeds. Instead of a fixed class list, you type the object names you care about; a CLIP text encoder turns them into embeddings that can be folded into the network weights, so the deployed detector runs at ordinary YOLO speed and exports to ONNX or TFLite. That made it a natural fit for zero-shot detection on edge hardware, and Ultralytics, Roboflow and ComfyUI all integrated it. The official repository's last commit came in February 2025, shortly after the V2.1 weights, it has never tagged a release, and recent issues get no maintainer reply even though the README still claims active development; an August 2026 report that the commercial licensing contact no longer answers is itself unanswered. The code is tied to PyTorch 1.11-era MMYOLO and MMDetection versions that stopped moving long ago, so a native install is painful, and the Ultralytics package is the easiest way to run it today. Licensing is GPL-3.0, or AGPL-3.0 through Ultralytics. For new work, YOLOE extends the same idea with text, visual and prompt-free modes.