Skip to main content
RF-DETR is a real-time transformer architecture for object detection and instance segmentation developed by Roboflow. Built on a DINOv2 vision transformer backbone, RF-DETR delivers state-of-the-art accuracy-latency trade-offs on Microsoft COCO and RF100-VL, with model sizes ranging from Nano to 2XLarge.
RF-DETR requires Python >= 3.10. Install it in a compatible environment before proceeding.

​
Key features

  • State-of-the-art on COCO — RF-DETR achieves the best accuracy–latency trade-off among real-time object detection and instance segmentation models on both COCO and RF100-VL.
  • Multiple model sizes — Six sizes from Nano to 2XLarge let you match your latency and accuracy requirements.
  • Detection and segmentation — A single, consistent API covers both object detection and instance segmentation tasks.
  • DINOv2 backbone — A windowed DINOv2 vision transformer backbone powers all model sizes.
  • Fine-tuning support — Fine-tune any model size on your own dataset in a single model.train() call.

​
Get started

Quickstart

Run your first detection or segmentation model in minutes.

Run detection

Load and run pre-trained RF-DETR detection models on images and video.

Train a model

Fine-tune RF-DETR on your custom dataset.

API reference

Explore the full Python API surface.

​
Model sizes

RF-DETR provides detection and segmentation models across six sizes.
SizeDetection classCOCO AP50:95Latency (ms)License
NRFDETRNano48.42.3Apache 2.0
SRFDETRSmall53.03.5Apache 2.0
MRFDETRMedium54.74.4Apache 2.0
LRFDETRLarge56.56.8Apache 2.0
XLRFDETRXLarge58.611.5PML 1.0
2XLRFDETR2XLarge60.117.2PML 1.0
All latency numbers were measured on an NVIDIA T4 using TensorRT, FP16, and batch size 1.

​
Research paper

RF-DETR was presented at ICLR 2026. The architecture is described in the paper RF-DETR: Neural Architecture Search for Real-Time Detection Transformers (arXiv:2511.09554). 
@misc{rf-detr,
    title={RF-DETR: Neural Architecture Search for Real-Time Detection Transformers},
    author={Isaac Robinson and Peter Robicheaux and Matvei Popov and Deva Ramanan and Neehar Peri},
    year={2025},
    eprint={2511.09554},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url={https://arxiv.org/abs/2511.09554},
}

​
Licensing

Licensing is split by component:
  • Apache 2.0 — The rfdetr package, all code, and Apache-designated model weights (Nano through Large for detection; all segmentation sizes).
  • PML 1.0 — Plus components, including RF-DETR-XL and RF-DETR-2XL detection models. Requires pip install rfdetr[plus].

Build docs developers (and LLMs) love

Get started for freeTalk to us