RF-DETR is a real-time transformer architecture for object detection and instance segmentation developed by Roboflow. Released as the open-source `rfdetr` package under Apache 2.0 and accompanied by an ICLR 2026 paper, it has gathered 7,657 GitHub stars by delivering state-of-the-art accuracy-versus-latency trade-offs on Microsoft COCO and the RF100-VL benchmark — and by doing so under a permissive license, in direct contrast to the AGPL-licensed YOLO family that has dominated production object detection. ## A DETR Built for Real-Time Use DETR-style detectors brought the clean, anchor-free, end-to-end transformer formulation to object detection but historically lagged on speed. RF-DETR closes that gap. Built on a DINOv2 vision transformer backbone, it pairs the representational strength of a self-supervised foundation model with an architecture tuned for low latency, supporting both detection and instance segmentation through a single, consistent API. The result is a model family that competes with — and frequently beats — the latest YOLO releases on the accuracy/latency frontier while keeping the transformer's simpler training and deployment story. ## Benchmark Results Roboflow reports benchmarks on both Microsoft COCO and RF100-VL, with all latency numbers measured on an NVIDIA T4 using TensorRT, FP16, and batch size 1. Across the detection lineup, RF-DETR-N reaches 48.4 COCO AP at 2.3 ms, RF-DETR-S hits 53.0 AP at 3.5 ms, RF-DETR-M reaches 54.7 AP at 4.4 ms, and RF-DETR-L delivers 56.5 AP at 6.8 ms — consistently ahead of comparably sized YOLO11, YOLO26, LW-DETR, and D-FINE variants at similar latencies. The instance segmentation models follow the same pattern, with RF-DETR-Seg variants posting strong COCO mask AP at competitive latencies. On the RF100-VL benchmark, which spans 100 diverse real-world domains, RF-DETR's lead is particularly pronounced, signaling strong generalization beyond the COCO distribution. ## Detection and Segmentation in One API A notable design choice is that RF-DETR handles both object detection and instance segmentation through the same consistent interface. Teams that previously needed separate models and codebases for bounding boxes and masks can standardize on one architecture, one training recipe, and one inference path. This unification reduces the operational surface area of a computer-vision system and makes it simpler to move a project from detection to segmentation as requirements evolve, without re-tooling the pipeline. ## Designed for Fine-Tuning RF-DETR is explicitly built to be fine-tuned on custom datasets, which is where the DINOv2 backbone pays off. The pretrained representations transfer well to specialized domains — manufacturing defects, medical imagery, retail, agriculture — that look nothing like COCO, and the strong RF100-VL results are the quantitative evidence for that transfer. Roboflow provides a Colab notebook for fine-tuning on a detection dataset and a Hugging Face Space for interactive trials, lowering the barrier for practitioners who want to adapt the model to their own data rather than train detection from scratch. ## Installation and Ecosystem The model installs as a single pip package — `pip install rfdetr` — in a Python 3.10+ environment, with a source install path available for accessing in-development features. RF-DETR slots naturally into Roboflow's broader open-source tooling, including the Supervision library for post-processing and visualization, and benchmarking is reproducible through Roboflow's single-artifact benchmarking repository. The combination of a one-line install, a published paper, public benchmarks, and an active Discord community has driven rapid adoption among computer-vision engineers. ## Licensing Nuance The core `rfdetr` package and the Apache-designated models — RF-DETR-N/S/M/L for detection and the full RF-DETR-Seg lineup for segmentation — are released under Apache 2.0, which permits unrestricted commercial use. The larger Plus components, including the RF-DETR-XL and 2XL detection models, ship as `rfdetr_plus` under the separate PML 1.0 license. For most production needs the Apache-2.0 models are more than sufficient, and the permissive licensing is a deliberate and meaningful differentiator from the AGPL-3.0 terms attached to many competing YOLO models, which can impose copyleft obligations on commercial deployments. ## Why It Matters With 7,657 stars, 977 forks, and an ICLR 2026 paper behind it, RF-DETR makes a strong case that the real-time detection crown no longer belongs by default to the YOLO lineage. By combining a DINOv2 foundation-model backbone, SOTA accuracy/latency trade-offs on COCO and RF100-VL, unified detection and segmentation, fine-tuning-first design, and genuinely permissive Apache-2.0 licensing for its core models, RF-DETR offers production computer-vision teams a compelling and license-friendly alternative for 2026.