Skip to content
Ultralytics YOLO26 logo

Ultralytics YOLO26

Current YOLO model family from Ultralytics: five sizes, NMS-free end-to-end detection plus segmentation, pose and oriented boxes, exportable to TensorRT, ONNX, OpenVINO and TFLite.

4/5 my assessment
Open-source
Artificial IntelligenceRobotics

Overview

YOLO26 launched on 14 January 2026 and heads the Ultralytics line that runs back through YOLO11, YOLOv8 and YOLOv5. The headline change is end-to-end NMS-free inference: one-to-one label assignment lets the network output final detections directly, so there is no non-maximum suppression step to reimplement in whatever runtime you deploy into. Distribution Focal Loss is gone too, which removes a regression head that historically broke ONNX and TensorRT exports in awkward ways. On the training side it adds ProgLoss for loss balancing, small-target-aware label assignment and a MuSGD optimizer blending SGD with Muon.

For a working engineer the value sits in the export path. One CLI call produces ONNX, TensorRT, CoreML, TFLite or OpenVINO artifacts, and the nano variant reports around 39.8 mAP at 640 px while running roughly 43 percent faster on CPU than YOLO11n. That combination is why it keeps turning up on Jetson boards, Raspberry Pi rigs and Intel-based line cameras. Five tasks ship out of the box, including oriented bounding boxes for aerial and web-inspection imagery.

The catch is legal, not technical. Everything is AGPL-3.0, and the network clause means serving a YOLO26 model from a closed application can obligate you to release that application's source. Ultralytics sells an Enterprise Licence that removes the obligation but publishes no price. Internal-only tooling is usually fine. Anyone building a product with vision inside it should either budget for the licence early or evaluate a permissively licensed detector such as RF-DETR before the architecture is locked.

Key Features

  • Released 14 January 2026 in five sizes from nano to extra large, covering detection, instance segmentation, classification, pose, oriented bounding boxes and tracking
  • End-to-end NMS-free inference through one-to-one label assignment, so the network emits final boxes with no post-processing step to port into your target runtime
  • Distribution Focal Loss removed from the regression head, which kills a long-standing source of ONNX and TensorRT export breakage
  • YOLO26n reports roughly 39.8 mAP at 640 px and runs about 43 percent faster on CPU than YOLO11n
  • ProgLoss progressive loss balancing plus small-target-aware label assignment, aimed squarely at small object accuracy
  • Single-command export to ONNX, TensorRT, CoreML, TFLite and OpenVINO

Where it holds

  • The export path holds up on real targets: Jetson, Raspberry Pi and Intel CPU cameras, from one CLI call
  • Dropping NMS removes the most common cause of works-in-Python, breaks-in-C++ detection bugs
  • Documentation, tutorials and hardware-vendor support are far ahead of every alternative detector, which matters when you hand a model to a plant engineer

Where it breaks

  • AGPL-3.0 by default. Serving a YOLO26 model from a closed application can pull that application's source into the copyleft obligation, and Ultralytics markets commercial use as requiring a paid licence
  • Enterprise licence pricing is unpublished and negotiated per deal, so budgeting starts with a sales call rather than a spreadsheet
  • The Pro tier at $29 per seat plus $30 in credits is a platform subscription, not a licence fix, and people confuse the two
  • Headline benchmarks are vendor-run on COCO. Independent reproductions of the 43 percent CPU claim are still thin

My Take

The licence is the first thing to settle, not the architecture. AGPL-3.0 covers YOLO26 exactly as it covered YOLOv8 and YOLO11, and the network clause bites the moment detections leave a proprietary service, which is how most vision ships. Set that aside and the January 2026 release is a real engineering improvement: NMS-free output and the removal of Distribution Focal Loss kill two of the nastiest export failure modes, and the nano model is roughly 43 percent faster on CPU than YOLO11n at similar accuracy. Roboflow's RF-DETR beats it on COCO AP and on licensing, but nothing else has this much tooling and hardware support behind it.

Francis Okafor
Francis Okafor AI & Tech Lead · Engineer

Quick Info

Pricing:
open-source
Openness:
Open source
Licence:
AGPL-3.0
Starting at:
Model code and weights are AGPL-3.0 and free to download. The Ultralytics Platform has a Free tier with $25 in signup credits, 100 GB storage, 100 models and 3 concurrent cloud trainings, and a Pro tier at $29 per seat per month plus $30 per seat per month in compute credits, with 500 GB storage and B200 or B300 GPU access. The commercial Enterprise Licence that removes the AGPL obligation is negotiated privately with no public price as of August 2026. Verified on ultralytics.com/pricing and ultralytics.com/license.
Added:
Aug 2026
Updated:
Aug 2026

Use Cases

computer vision manufacturing robotics

Judge it on your own work

The notes above say where Ultralytics YOLO26 holds and where it breaks. The fastest check is your own workload.

Visit website ↗

Alternatives to Ultralytics YOLO26