YOLO26, released in January 2026 by Ultralytics, drops Non‑Maximum Suppression and the Distribution Focal Loss module, slashing latency while keeping accuracy. The family spans Nano to Extra Large, supporting detection, segmentation, pose estimation, classification, and oriented object detection.

Edge deployment gains drive the design. Removing NMS lets the model output predictions directly, cutting inference time by up to 43% on CPUs compared to its predecessor. The loss‑module removal also widens compatibility with TFLite, Core ML, Open VINO, and TensorRT, making YOLO26 suitable for low‑power devices.

Benchmarking on COCO shows the Nano variant achieves 40.9 mAP at 938 ms on CPUs, while the Extra Large reaches 57.6 mAP with 114 ms on a T4 TensorRT. These figures outpace older YOLO releases and match the performance of RF‑DET R, LW‑DET R, and D‑FINE.

For developers, YOLO26 offers a unified training and export pipeline through Roboflow, enabling quick iteration from labeling to deployment. The model’s compact architecture and broad hardware support position it as a practical choice for real‑time vision on edge devices.