RT-DETR

weights

open(RT-DETRv4 S to X checkpoints linked from its README

data

documented-not-released(COCO 2017 is public, but v4 distills from a DINOv3 ViT-B/16 teacher trained on the unreleased LVD-1689M corpus)

code

open(training code and configs for v1, v2 and v4)

license

Apache-2.0(repository LICENSE for v1, v2 and v4

Every version is Apache-2.0, code and weights, with the full training code and configs published. The training data cannot be fully reproduced, though: RT-DETRv4's recipe distills features from a DINOv3 ViT-B/16 teacher, and Meta has not released that teacher's LVD-1689M training images, so COCO being public does not make the detector's full training data available. The teacher itself must be obtained separately under Meta's DINOv3 License.

• https://huggingface.co/api/models?author=PekingU&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27

  Eleven PekingU RT-DETR and RT-DETRv2 checkpoints, all gated false and tagged apache-2.0, including the _coco_o365 variants.

• https://raw.githubusercontent.com/facebookresearch/dinov3/main/MODEL_CARD.md recorded 2026-09-27

  DINOv3 model card: "Web dataset (LVD-1689M): a curated dataset of 1,689 millions of images extracted from a large data pool of 17 billions web images collected from public posts on Instagram"; no release of the dataset.

• https://raw.githubusercontent.com/lyuwenyu/RT-DETR/main/LICENSE recorded 2026-09-27

  RT-DETR (v1 and v2) LICENSE is the Apache License, Version 2.0.

• https://raw.githubusercontent.com/lyuwenyu/RT-DETR/main/README.md recorded 2026-09-27

  README: "Release the **newest** member of the RT-DETR family: RT-DETRv4: Painlessly Furthering Real-Time Object Detection with Vision Foundation Models" at github.com/RT-DETRs/RT-DETRv4.

• https://raw.githubusercontent.com/RT-DETRs/RT-DETRv4/main/LICENSE recorded 2026-09-27

  RT-DETRv4 LICENSE is the Apache License, Version 2.0.

• https://raw.githubusercontent.com/RT-DETRs/RT-DETRv4/main/README.md recorded 2026-09-27

  RT-DETRv4 README: model table RT-DETRv4-S/M/L/X (COCO AP 49.8 to 57.0) each with a ckpt download link; COCO2017 dataset preparation; training command "torchrun ... train.py -c configs/rtv4/rtv4_hgnetv2_${model}_coco.yml"; "We use the ViT-B/16-LVD-1689M model from DINOv3" as the teacher.

• https://ungh.cc/repos/RT-DETRs/RT-DETRv4/files/main recorded 2026-09-27

  RT-DETRv4 repository tree includes train.py and the configs/rtv4 training configs.

Hugging Face downloads over the trailing 30 days for the two most-used checkpoints, which are the ones declared; the other PekingU checkpoints add comparatively little. RT-DETRv4 checkpoints are downloaded from a file-sharing link that publishes no counts.

• https://huggingface.co/api/models?author=PekingU&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27

  rtdetr_r101vd_coco_o365 896,124 and rtdetr_v2_r18vd 526,281 downloads in the trailing 30 days (1,422,405 together); the eleven checkpoints sum to 2,065,416.

RT-DETR and RF-DETR are the same kind of product, real-time transformer detectors pretrained on COCO's 80 classes, so detecting anything else still takes a training run on the user's own labels. RT-DETRv4 reaches 57.0 COCO AP at its largest size.

• https://raw.githubusercontent.com/lyuwenyu/RT-DETR/main/rtdetrv2_pytorch/README.md recorded 2026-09-27

  RT-DETRv2 README lists COCO and Objects365-pretrained checkpoints and the torchrun tools/train.py training command.

• https://raw.githubusercontent.com/RT-DETRs/RT-DETRv4/main/README.md recorded 2026-09-27

  RT-DETRv4 README: "RT-DETRv4 is the new version of the state-of-the-art real-time object detector family, RT-DETR."; RT-DETRv4-S 49.8, M 53.7, L 55.4, X 57.0 COCO AP; "To train on your custom dataset, you need to organize it in the COCO format."