RF-DETR
Real-time detection without the AGPL question
What it is
Roboflow's real-time transformer architecture for detection and instance segmentation, built on a DINOv2 backbone and reported state-of-the-art on COCO and RF100-VL. Six detection sizes from Nano to 2XLarge, optimized for fine-tuning on custom datasets. Accepted at ICLR 2026.
Why it's interesting
A genuinely permissive, high-accuracy real-time detector — the strong Apache-2.0 alternative to the AGPL-licensed YOLO ecosystem for commercial use, with edge-friendly latency (2-17ms on a T4).
Use cases
- Real-time detection on custom datasets
- Instance segmentation
- Edge and embedded vision
Who it's for
CV engineers, product teams needing commercial-friendly detectors
Setup
Easy. Python 3.10+, pip install rfdetr; a T4-class GPU or better for real-time
Limitations & cautions
Segmentation is instance-level only, and published latency figures assume TensorRT with FP16.
Editorial takeaway
The detector to reach for when legal asks 'what's our YOLO exposure?' and you'd rather not have that meeting.