Ultralytics YOLO26

YOLO26: A better, faster, smaller YOLO model

Ultralytics on 25 Sep 2025 shared more information about YOLO26

Here’s an overview of the computer vision tasks supported by YOLO26:

• Object detection: YOLO26 can identify and locate multiple objects within an image or video frame.
  ‍
• Instance segmentation: Going a step beyond detection, YOLO26 can generate pixel-perfect boundaries around each object it identifies. 
  ‍
• Image classification: The model can analyze an entire image and assign it to a specific category or label.
  
  
• Pose estimation: YOLO26 can detect keypoints and estimate poses for humans as well as other objects.
  
  
• Oriented bounding boxes (OBB): The model can detect objects at any angle, which is especially useful for aerial, drone, and satellite imagery, where items like buildings, vehicles, or crops may not be aligned with the image frame.
  
  
• Object tracking: YOLO26 can be used to track objects across video frames or real-time streams.

In fact, the smallest version of YOLO26, the nano model, now runs up to 43% faster on standard CPUs, making it especially well-suited for mobile apps, smart cameras, and other edge devices where speed and efficiency are critical.

Here’s a quick recap of YOLO26’s features and what users can look forward to:

• DFL removal: We removed the Distribution Focal Loss module from the model’s architecture. Regardless of the object sizes in an image, YOLO26 can place tailored bounding boxes while running more efficiently.
  ‍
• End-to-end NMS-free inference: YOLO26 adds an optional mode that doesn’t need Non-Maximum Suppression (NMS), a step normally used to remove duplicate predictions, making deployment simpler and faster for real-time use.
  ‍
• ProgLoss and STAL: These improvements make training more stable and significantly boost accuracy, especially for detecting small objects in complex scenes.
  ‍
• MuSGD optimizer: YOLO26 uses a new optimizer that combines the strengths of two training optimizers (Muon and SGD), helping the model learn faster and reach higher accuracy.

Simplifying deployment with Ultralytics YOLO26 

Whether you are working on mobile apps, smart cameras, or enterprise systems, deploying YOLO26 is simple and flexible. The Ultralytics Python package supports a constantly growing number of export formats, which makes it easy to integrate YOLO26 into existing workflows and makes it compatible with almost any platform. 

A few of the export options include TensorRT for maximum GPU acceleration, ONNX for broad compatibility, CoreML for native iOS apps, TFLite for Android and edge devices, and OpenVINO for optimized performance on Intel hardware. This flexibility makes it straightforward to take YOLO26 from development to production without extra hurdles.

Another crucial part of deployment is making sure models run efficiently on devices with limited resources. This is where quantization comes in. Thanks to its simplified architecture, YOLO26 handles this exceptionally well.  It supports INT8 deployment (using 8-bit compression to reduce size and improve speed with minimal accuracy loss) as well as half-precision (FP16) for faster inference on supported hardware. 

Most importantly, YOLO26 delivers consistent performance across these quantization levels, so you can rely on it whether it’s running on a powerful server or a compact edge device.

Documentation:

https://docs.ultralytics.com/models/yolo26/

 Please read more on Ultralytics blog

https://www.ultralytics.com/blog/meet-ultralytics-yolo26-a-better-faster-smaller-yolo-model