RF-DETR

weights

open(Nano to Large and all segmentation sizes on the Hub)

data

documented-not-released(Objects365 pseudo-labeled with SAM2, then COCO

code

partial(fine-tuning code

license

Apache-2.0(code and Nano-Large weights)+PML-1.0(Roboflow Platform Model License on the XL and 2XL detection weights)

Most of the line is Apache-2.0, code and weights alike. The two largest detection models are the exception: they ship in a separate rfdetr_plus package under Roboflow's Platform Model License, which allows use only under a Roboflow platform plan or other agreement and an account in good standing. The map treats the most restrictive license on any distributed checkpoint as the line's, so that platform license governs here. PML-1.0 does not permit commercial use, so the license tier is commercial_forbidden.

• https://arxiv.org/html/2511.09554v2 recorded 2026-09-26

  Paper: "First, we pseudo-label Objects-365 ... with SAM2 ... to allow us to pre-train the segmentation and detection heads on the same data." No release of the pseudo-labeled set is stated.

• https://huggingface.co/api/models?author=Roboflow&sort=downloads&limit=50&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-26

  Roboflow's Hub listing: every RF-DETR checkpoint (base, nano, small, medium, large and the seg sizes) tagged apache-2.0 and not gated.

• https://raw.githubusercontent.com/roboflow/rf-detr/HEAD/docs/learn/train/index.md recorded 2026-09-26

  Training docs: "Fine-tune from COCO-pretrained checkpoints (Nano to 2XLarge) for fastest convergence".

• https://raw.githubusercontent.com/roboflow/rf-detr/HEAD/LICENSE recorded 2026-09-26

  LICENSE is the Apache License, Version 2.0, "Copyright 2025 Roboflow, Inc."

• https://raw.githubusercontent.com/roboflow/rf-detr/HEAD/README.md recorded 2026-09-26

  README License section: "The open-source rfdetr package and Apache-designated model weights are licensed under Apache License 2.0." / "Plus components, including the rfdetr_plus extension and RF-DETR-XL / RF-DETR-2XL detection models, are licensed under PML 1.0."

• https://roboflow.com/licensing recorded 2026-09-26

  Roboflow licensing table: "RF-DETR (XL and 2XL) | Object Detection | PML 1.0"; "RF-DETR (Nano to Large) | Object Detection | Apache-2.0".

• https://roboflow.com/platform-model-license-1-0 recorded 2026-09-26

  Platform Model License 1.0: "You may only use, reproduce, modify, or distribute the Software if you ... have accepted and are complying with the terms (including pricing, usage limits, and other usage metrics) of a separate platform plan".

• https://ungh.cc/repos/roboflow/rf-detr/files/develop recorded 2026-09-26

  Repository tree at develop includes src/rfdetr/training/ (trainer.py, cli.py) and a train-coco2017 cookbook.

Measured on monthly PyPI downloads of rfdetr, the documented way to install and run the models; Hugging Face counts on the checkpoints are far lower because the package fetches weights itself.

• https://pypistats.org/api/packages/rfdetr/recent recorded 2026-09-26

  last_month 206,471 downloads of rfdetr

RF-DETR's 2XL size is described by its authors as the first real-time detector above 60 AP on COCO, and the line ships detection, segmentation and keypoint heads. Detecting anything beyond COCO's 80 classes still takes a fine-tuning run. It has no open-vocabulary model of the kind Ultralytics bundles, so it sits with the toolkits rather than with the models that work on a new task untrained.

• https://arxiv.org/abs/2511.09554 recorded 2026-09-26

  Abstract: "RF-DETR (2x-large) is the first real-time detector to surpass 60 AP on COCO."

• https://huggingface.co/Roboflow/rf-detr-base/raw/main/README.md recorded 2026-09-26

  Card: "These checkpoints are trained on the standard COCO 2017 object detection dataset label space (80 categories)".

• https://raw.githubusercontent.com/roboflow/rf-detr/HEAD/README.md recorded 2026-09-26

  README detection table (COCO AP50:95 / RF100VL AP50:95): N 48.4/57.7, S 53.0/60.2, M 54.7/61.2, L 56.5/62.2, XL 58.6/62.9, 2XL 60.1/63.2; segmentation and keypoint (preview) tables alongside.