RF-DETR, YOLO26: real-time SOTA vision hits mobile

On-device vision has finally reached the "zero-latency" threshold for developers, moving beyond basic detection into high-fidelity instance segmentation. RF-DETR Nano's use of a DINOv2 transformer backbone achieves 48.0 AP on COCO, a massive leap over previous generation "Nano" models, while YOLO26 introduces MuSGD optimization and 43% faster CPU inference for edge devices. NMS-free architectures are the critical breakthrough, removing the most significant computational bottleneck in mobile vision deployment. Additionally, the availability of RF-DETR under an Apache 2.0 license provides a permissive commercial path compared to the more restrictive AGPL licenses common in the YOLO lineage, and compatibility with CoreML, TFLite, and ONNX ensures these models are production-ready for cross-platform mobile development.