Get Started

YOLOv5 vs YOLO26

The evolution of computer vision has been defined by the continuous push for faster, more accurate, and more accessible models. Ultralytics YOLOv5 and Ultralytics YOLO26 mark two points in that progression: a widely deployed legacy model and the current recommended Ultralytics model.

This guide provides a comprehensive technical breakdown of both architectures, highlighting their performance metrics, structural differences, and ideal deployment scenarios.

Model Overviews#

YOLOv5: The Industry Workhorse#

Released in 2020, YOLOv5 made object detection far more accessible. By implementing the architecture natively in the PyTorch framework, it gave developers a simple path from installation to training and deployment.

YOLOv5 established the foundation for the actively maintained Ultralytics ecosystem. It introduced strong data augmentation techniques, efficient training loops, and highly optimized export paths to edge formats like CoreML and ONNX. Its ease of use and low memory requirements during training made it a staple for startups and researchers worldwide.

YOLO26: The Next-Gen Vision AI Standard#

Released in January 2026, Ultralytics YOLO26 is the latest generation of Ultralytics YOLO. It integrates lessons learned from intervening generations like YOLOv8 and YOLO11, while introducing training techniques inspired by Large Language Model (LLM) training.

YOLO26 offers a strong balance of accuracy and speed and was designed specifically for edge computing scenarios.

Other Ultralytics Models

If you are migrating an older codebase, you might also be interested in comparing YOLOv5 with YOLO11, the previous generation model supporting diverse tasks like pose estimation and oriented bounding boxes (OBB).

Architectural Innovations in YOLO26#

While YOLOv5 relies on anchor-based detection heads and standard loss functions, YOLO26 reworks the head, loss functions, and optimizer to remove deployment bottlenecks.

  1. End-to-End NMS-Free Design: The most significant difference is YOLO26's native end-to-end inference option. Unlike YOLOv5, which requires Non-Maximum Suppression (NMS) to filter redundant bounding boxes, YOLO26's optional one-to-one head (nms=False) skips this post-processing step. This keeps inference latency more predictable and simplifies integration into C++ or embedded hardware.
  2. DFL Removal: YOLO26 removes Distribution Focal Loss (DFL). This architectural choice simplifies model export and improves compatibility with low-power edge devices and accelerators that often struggle with complex operators.
  3. MuSGD Optimizer: Taking cues from Moonshot AI's Kimi K2, YOLO26 utilizes the MuSGD Optimizer, a hybrid of SGD and Muon. This brings the stability and fast convergence seen in LLM training to computer vision.
  4. ProgLoss + STAL: YOLO26 uses ProgLoss and STAL to improve detection of small and dense objects, a historical challenge for YOLOv5.

Performance Comparison#

When comparing the models on the COCO dataset, YOLO26 shows higher accuracy (mAP) at every scale while using fewer parameters from the medium variant upward.

Modelsize
(pixels)
mAPval50-95Speed
CPU ONNX
(ms)
Speed
T4 TensorRT10
(ms)
params
(M)
FLOPs
(B)
YOLOv5n64028.0-1.121.94.5
YOLOv5s64037.4-1.927.216.5
YOLOv5m64045.4-4.0321.249.0
YOLOv5l64049.0-6.6146.5109.1
YOLOv5x64050.7-11.8986.7205.7
YOLO26n64040.938.91.72.45.5
YOLO26s64048.687.22.59.520.9
YOLO26m64053.1220.04.720.468.4
YOLO26l64055.0286.26.224.886.8
YOLO26x64057.5525.811.855.7194.4

Note: The YOLO26 Nano (YOLO26n) achieves 40.9 mAP compared to YOLOv5n's 28.0 mAP, a 12.9-point gain for only 0.5M additional parameters, while DFL removal and the NMS-free head simplify CPU and edge deployment.

Versatility and Task Support#

YOLOv5 is primarily known for object detection. While later updates introduced basic segmentation, YOLO26 was built from the ground up to be a unified multi-task engine.

YOLO26 inherently supports:

  • Instance Segmentation: Featuring task-specific multi-scale protos and semantic segmentation loss.
  • Semantic Segmentation: Dense per-pixel class labeling.
  • Depth Estimation: Per-pixel depth prediction from a single image.
  • Image Classification: Standard full-image categorization.
  • Pose Estimation: Utilizing Residual Log-Likelihood Estimation (RLE) for more accurate keypoint localization.
  • Oriented Bounding Boxes (OBB): Including specialized angle loss to resolve boundary discontinuity issues, critical for satellite image analysis.
Ecosystem Integration

Both models benefit from the Ultralytics Platform, which provides data annotation, cloud training, and one-click deployment.

Usage and Code Examples#

The Ultralytics Python API makes switching between models simple. Because both models share the same well-maintained ecosystem, updating a YOLOv5u pipeline in the ultralytics package to YOLO26 only requires changing the weights file.

Python Example#

from ultralytics import YOLO

# To use YOLOv5, load a v5 weights file
# model = YOLO("yolov5su.pt")

# Migrate to the recommended YOLO26 model
model = YOLO("yolo26n.pt")

# Train on the COCO8 dataset; optimizer="auto" selects MuSGD for longer runs (>10k iterations)
results = model.train(
    data="coco8.yaml",
    epochs=100,
    imgsz=640,
    batch=32,  # YOLO26's low memory footprint allows larger batch sizes
)

# Run an NMS-free inference
predictions = model("https://ultralytics.com/images/bus.jpg", nms=False)
predictions[0].show()

CLI Example#

You can deploy YOLO26 directly via the command line using the TensorRT integration for maximum GPU throughput:

# Export the model to TensorRT format
yolo export model=yolo26n.pt format=engine

# Run inference with the compiled engine
yolo predict model=yolo26n.engine source=path/to/video.mp4

Ideal Use Cases#

When to choose YOLO26#

For new computer vision projects, YOLO26 is the recommended choice.

  • Edge AI and IoT: Its up to 43% faster CPU inference (YOLO26n vs YOLO11n) and removal of DFL make it well suited to deployment on a Raspberry Pi or mobile devices.
  • High-Speed Pipelines: The optional NMS-free head provides more predictable latency, which matters for autonomous robotics and real-time security alarm systems.
  • Complex Scenarios: If your application requires detecting small objects (e.g., drone monitoring) or rotated objects (OBB), YOLO26's training improvements (ProgLoss + STAL, plus a dedicated angle loss for OBB) improve accuracy.

When to choose YOLOv5#

  • Legacy Systems: If your production environment has hardcoded dependencies on YOLOv5's specific anchor generation or NMS parsing logic, migrating might require a brief refactoring period.
  • Specific Academic Baselines: Researchers often use YOLOv5 as a classic baseline to demonstrate the historical progression of object detection architectures.

Summary#

The transition from YOLOv5 to YOLO26 is more than an iterative update: it changes how object detection models are trained and deployed. With the MuSGD optimizer, an optional NMS-free head that drops post-processing, and faster CPU inference, Ultralytics YOLO26 delivers a strong balance of speed and accuracy.

While YOLOv5 remains the model that made vision AI accessible to a broad developer community, developers building new production applications should start with YOLO26.

Comments

Ask a question, share your results, or help someone else.