Ultralytics YOLO27#

Overview#

Ultralytics YOLO27 is a family of real-time vision models whose detection models use two complementary designs: a streamlined CNN architecture for the compact N and S models, and a query-based, NMS-free architecture for the larger M and L models. Both designs are end-to-end and deploy through the same interface.

YOLO27 is the upcoming model family in the Ultralytics YOLO series, succeeding YOLO26. It comes in four sizes — N, S, M and L — and supports object detection, instance segmentation, semantic segmentation, depth estimation, classification, pose estimation and oriented object detection. The two-design split applies to detection; all other tasks use the CNN architecture.

Across its four detection scales, YOLO27 reaches 42.3-60.4 mAP on COCO at 0.62-2.32 ms latency on an NVIDIA RTX PRO 6000 — and up to 61.2 mAP with YOLO27l at a larger 800-pixel input. YOLO27l is the first Ultralytics model to surpass 60 mAP on COCO, while the compact YOLO27n/s improve on YOLO26n/s accuracy at essentially the same speed.

from ultralytics import YOLO

model = YOLO("yolo27n.pt")  # load a pretrained YOLO27n model
results = model("path/to/bus.jpg")  # run inference

Key Features#

• Dual-scale detection Standard detectors predict objects on three feature maps — fine, medium, and coarse. YOLO27 N and S drop the medium one and predict only on a fine map (for small objects) and a coarse map (for large objects), with a fixed scaling on the fused features keeping the two scales balanced. This removes a large chunk of detection-head computation, making the models faster while training as reliably as the full three-scale design.

• Stronger small-object detection The early, high-resolution feature stage is widened so it can capture more fine-grained detail. Combined with the surviving fine prediction map, this improves localization and regression for small objects — the hardest category for compact models.

• Foreground alignment supervision During training, an extra lightweight branch learns to tell "object" from "background" at every location. It is designed to close the gap between the denser one-to-many supervision used during training and the one-to-one head that produces the final predictions — cutting that accuracy gap from 0.9/0.8 mAP on YOLO26n/s to just 0.4 mAP on YOLO27n/s, so the deployed one-to-one head keeps nearly all of the training-time accuracy. The branch is used only during training and is removed for inference and export, so it costs nothing at deployment.

• Query-based detection without NMS The larger models replace dense prediction with a transformer decoder that refines a fixed set of object queries and directly outputs the final detections — no non-maximum suppression post-processing needed. YOLO27m pairs this decoder with the proven YOLO26-style convolutional backbone, while YOLO27l keeps the same FPN/PAN neck and swaps in an UltraViT backbone that uses self-attention in its deepest stage to capture global context.

• One simple interface Both architectures are used through the same YOLO Python interface. The right training, validation, prediction, and export pipeline is selected automatically from the model, so code written for one YOLO27 scale works unchanged for the others.

Which YOLO27 Should I Use?#

The following guidance is for choosing a model once YOLO27 is released. For projects today, use YOLO26.

• YOLO27n / YOLO27s — edge devices, drones, and real-time video: the fastest models in the family, with improved small-object detection from the dual-scale design. See NVIDIA Jetson and Raspberry Pi for device-specific deployment.
• YOLO27m — the accuracy-speed sweet spot on GPUs: improves on YOLO26m by 3.2 mAP at the same latency, making it the default choice for production GPU deployment.
• YOLO27l — accuracy-critical applications: the first Ultralytics model above 60 mAP on COCO, reaching 61.2 mAP at a larger input size while staying real-time on GPU.

To compare against other families, see all Ultralytics models. To measure the accuracy-speed trade-off on your own hardware rather than ours, see Benchmark mode.

Supported Tasks and Modes#

The following table previews planned task and mode support across the four YOLO27 model scales. Checkmarks indicate intended support at release, not current public availability; all listed model files are unreleased.

YOLO27 Detect, Segment, Pose, and OBB models also work with Track mode, which runs on top of predict for multi-object tracking across video frames.

Performance Metrics#

Detection accuracy is reported on the COCO validation set. See YOLO Performance Metrics for explanations of mAP, precision, and recall. Inference speed is measured on an NVIDIA RTX PRO 6000 (TensorRT 11, FP16) for GPU and an AMD EPYC 9655 (ONNX Runtime, FP32) for CPU. After release, accuracy numbers can be reproduced with yolo val model=yolo27n.pt data=coco.yaml.

Params and FLOPs values are for the fused model after Conv/BatchNorm folding and removal of the unused detection branch. Speed measurements select the NMS-free head with nms=False. Pretrained checkpoints retain the full training architecture and may show higher counts.

Usage Examples#

This section provides simple YOLO27 training and inference examples. For full documentation on these and other modes, see the Predict, Train, Val, and Export docs pages. For custom datasets, see Tips for Best Training Results and the Hyperparameter Tuning Guide.

Note that the example below is for YOLO27 Detect models for object detection. For additional supported tasks, see the Segment and Classify docs.

from ultralytics import YOLO

# Load a COCO-pretrained YOLO27n model
model = YOLO("yolo27n.pt")

# Run inference with the YOLO27n model on the 'bus.jpg' image
results = model("path/to/bus.jpg")

# Train the model on the COCO8 example dataset for 100 epochs
results = model.train(data="coco8.yaml", epochs=100, imgsz=640)

Only this documentation preview is available now. YOLO27 model weights, configurations, and implementation code remain unreleased. Release instructions will be added to this page when the models become available.

FAQ#