YOLO27 Coming Soon: What Improves over YOLO26 and YOLOv8

Highlights

On the COCO validation set at 640 pixels, mAP50-95 for the four detection models is 42.3, 49.6, 55.8, and 60.4. YOLO27l is the first Ultralytics model above 60 mAP on COCO. At an 800-pixel input it reaches 61.2, with a latency of 2.9 ms. With speed essentially level, N and S are more accurate than YOLO26n/s.

Beyond detection, the plan includes instance segmentation, semantic segmentation, depth estimation, classification, pose estimation, and oriented object detection. The two architectures are for detection only. The other tasks all use a CNN. Detect, Segment, Pose, and OBB are also planned to support video tracking.

Where the accuracy differs from YOLO26 and YOLOv8

All three generations report mAP50-95 on COCO val2017, single model, single scale, 640 pixels, so that column can be compared directly. YOLO26 has a separate end-to-end column (e2e, NMS off at inference), usually 0.6 to 0.8 below the main score. The YOLO27 preview table does not split out that column, and it says N/S accuracy depends on the detection head that is selected. The YOLO26 numbers below use the main scores from its documentation.

Latency cannot be compared across the table. The YOLOv8 GPU numbers come from A100 TensorRT, YOLO26 from T4 TensorRT 10, and YOLO27 from RTX PRO 6000 TensorRT 11. The devices differ by an order of magnitude, so the milliseconds cannot be used to rank the three generations by speed. The comparison below is accuracy, parameter count, and compute only.

Read each size on its own. The accuracy gain for N and S is about 1 to 1.4 points. Compute is higher than YOLO26n/s, and still below or close to the same-size FLOPs of YOLOv8. M is the more notable case: 22.8M parameters and 65.0B FLOPs, below YOLOv8m and also below the compute of YOLO26m, while mAP reaches 55.8. L is not a free upgrade at the same size. YOLO27l has 72.3M parameters and 165.3B FLOPs, close in bulk to YOLOv8l and clearly larger than YOLO26l at 24.8M / 86.8B. What that buys is 60.4 mAP, which is also above 57.5 for YOLO26x and 53.9 for YOLOv8x.

Five architectural changes

How to choose a size after release

Projects already underway should keep using YOLO26. The choices below apply after YOLO27 weights are public.

• YOLO27n / YOLO27s: edge devices, drones, and real-time video. The two fastest sizes in the family, with a two-scale design that strengthens small objects. The official deployment notes will cover NVIDIA Jetson and Raspberry Pi.
• YOLO27m: the default production size for accuracy and speed on a GPU. Accuracy is higher than YOLO26m at a similar latency.
• YOLO27l: accuracy first. 60.4 mAP and 2.32 ms at 640 pixels; 61.2 mAP and 2.9 ms at 800 pixels, still real time on an RTX PRO 6000.

Zhcloudlink cameras and edge analytics boxes should still ship on released models. Do not write preview mAP into a solution commitment before YOLO27 weights and on-device measurements exist. For how chip compute and vision models fit together, see the RK3588 / RK3576 / RK3572 specification comparison.

Tasks and modes planned for support

A mark means support is planned at release. It does not mean the weights can be downloaded now. None of the weight files in the table have been released.

Performance figures

GPU speed was measured on an NVIDIA RTX PRO 6000 with TensorRT 11 and FP16. CPU speed was measured on an AMD EPYC 9655 with ONNX Runtime and FP32. After release, detection accuracy can be reproduced with yolo val model=yolo27n.pt data=coco.yaml. Parameter count and FLOPs are for the fused model: Conv and BatchNorm are folded, and unused detection branches are removed. Detection speed uses nms=False. A pretrained checkpoint keeps the full training structure, so parameter count and FLOPs may be higher.

Detection · COCO, 80 classes

Instance segmentation · COCO, 80 classes, 640 pixels

Semantic segmentation · Cityscapes, 19 classes, 1024×2048

Depth estimation · multi-dataset pretraining, 768 pixels

Classification · ImageNet, 1000 classes, 224 pixels

Pose · COCO, class person, 640 pixels

Oriented detection · DOTAv1, 15 classes, 1024 pixels

Usage after release

The examples apply only after release. Pass a COCO-pretrained .pt or a config .yaml to YOLO(). Training tips and hyperparameter search for custom data follow the Ultralytics documentation at release.

from ultralytics import YOLO

model = YOLO("yolo27n.pt")
results = model("path/to/bus.jpg")
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

yolo predict model=yolo27n.pt source=path/to/bus.jpg

yolo train model=yolo27n.pt data=coco8.yaml epochs=100 imgsz=640

Sources for comparison: Ultralytics YOLO26 documentation, Ultralytics YOLOv8 documentation. The YOLO27 numbers come from its official introduction, which is not final.