RF-DETR is a real-time object detection and instance segmentation architecture developed by Roboflow and published as an open-source project on GitHub, where it has earned close to 10,000 stars. Built on transformer detection technology, it sets state-of-the-art results on the COCO benchmark while staying fast enough for live video.
RF-DETR combines a DINOv2 vision backbone with a DETR-style detection head, and it is designed from the start for fine-tuning rather than only zero-shot inference. You bring a labeled dataset, train for a modest number of epochs, and get a compact model that runs in real time on edge or server hardware. The project ships detection and segmentation variants in several sizes, with code accepted at ICLR 2026, and it integrates tightly with the Roboflow platform for dataset management, training and deployment.
The RF-DETR code is open source under the Apache License 2.0 and free to use. Roboflow’s hosted platform, which provides dataset tooling, cloud training and managed deployment around it, offers a free tier with paid plans for larger workloads, so the overall offering is best described as Freemium.
RF-DETR is for computer vision engineers and product teams who need accurate detection or segmentation on their own classes without spending weeks tuning a model zoo. It is a strong fit for manufacturing inspection, retail analytics, robotics and any application where a fine-tuned, fast transformer detector beats a general-purpose zero-shot model.