By UltralyticsLearning Objectives
- Understand what YOLO26 is and what makes the YOLO family a default choice for real-time vision
- Identify the vision tasks a single YOLO26 model can perform
- Evaluate where YOLO26 fits — and where it does not — in a production vision stack
What Is YOLO26?
Ultralytics YOLO26 is the 2026 generation of YOLO ("You Only Look Once"), the most widely used family of real-time computer-vision models. Introduced in a June 2026 paper, "Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models," it folds five vision tasks into one architecture: object detection, instance segmentation, pose estimation, image classification, and oriented (rotated) box detection. It ships in five sizes — from nano to extra-large — so the same model family scales from a tiny edge device to a server GPU.
YOLO's appeal has always been speed plus simplicity: it is fast enough to run on live video, small enough to deploy at the edge, and easy to train on your own images. YOLO26 pushes on all three.
💡Key Concept
Real-time vision, in plain terms. A real-time detector looks at each video frame and reports what objects are present and where — fast enough to keep up with the camera. That is the technology behind everything from factory-line inspection and retail analytics to drones, sports tracking, and driver-assistance features.
What's New in YOLO26
- End-to-end inference — YOLO26 uses a dual-head design that removes the separate non-maximum-suppression cleanup step older detectors needed, simplifying deployment and lowering latency.
- Faster on CPUs — the smallest model runs up to 43 percent faster on a standard CPU than the equivalent YOLO11 model, which matters when there is no GPU at the edge.
- A better accuracy-speed trade-off — across its five sizes, YOLO26 advances the frontier of "how accurate can you be at a given speed" on the standard COCO benchmark.
- Open-vocabulary option (YOLOE-26) — an extension that can detect objects from a text prompt rather than only the fixed categories it was trained on.
- New training recipe — a hybrid optimizer adapted from large-language-model training, plus label-assignment tricks that improve detection of small objects.
Performance
On the standard COCO object-detection benchmark, YOLO26 reaches 40.9 to 57.5 mean average precision (mAP) across its five sizes, at roughly 1.7 to 11.8 milliseconds per image on an Nvidia T4 GPU — fast enough for real-time video. The open-vocabulary YOLOE-26 variant reaches 40.6 average precision on the LVIS benchmark using only text prompts.
| Capability | YOLO26 |
|---|---|
| Tasks | Detection, segmentation, pose, classification, oriented boxes |
| Model sizes | 5 (nano / small / medium / large / extra-large) |
| COCO accuracy | 40.9-57.5 mAP |
| Speed (T4 GPU) | 1.7-11.8 ms per image |
| CPU speedup vs YOLO11 | Up to 43% faster (nano) |
| Open vocabulary | Yes (YOLOE-26 extension) |
Pricing
- Full models and training code on GitHub
- Best for research, learning, and open projects
- Use in closed-source commercial products
- Removes AGPL copyleft obligations
- Contact Ultralytics
YOLO26 follows Ultralytics' long-standing dual-license model: free under the AGPL-3.0 open-source license, with a paid commercial license for companies that need to embed it in closed-source products.
Strengths
- One model, many tasks — detection, segmentation, pose, classification, and oriented boxes from a single architecture
- Fast everywhere — real-time on GPUs and meaningfully quicker on CPUs than the prior generation
- Easy to adopt — the Ultralytics toolkit makes training on your own images approachable, with a huge community and documentation base
- Open vocabulary — the YOLOE-26 extension detects objects from text prompts, not just fixed training categories
Limitations and Considerations
- License obligations — the free tier is AGPL-3.0; embedding it in a closed-source commercial product requires a paid Ultralytics license
- Not a multimodal chatbot — YOLO26 is a specialized vision model, not a general vision-language model that can reason or converse about images
- Benchmarks are not deployment — real-world accuracy depends heavily on training data quality and the specific cameras and conditions you deploy into
- Recent release — published June 2026; expect rapid point updates as the community puts it through production use
Company Details
| Detail | Info |
|---|---|
| Developer | Ultralytics |
| Released | June 2026 |
| Tasks | Detection, segmentation, pose, classification, oriented detection |
| License | AGPL-3.0 (open) + paid enterprise license |
| Availability | GitHub (ultralytics/ultralytics) |
Key Takeaways
- YOLO26 is the 2026 generation of the popular YOLO real-time vision family — one model that handles detection, segmentation, pose, classification, and oriented boxes across five sizes
- It runs end-to-end without the old non-maximum-suppression step, and its smallest model is up to 43 percent faster on CPUs than YOLO11
- An open-vocabulary extension, YOLOE-26, can detect objects from text prompts rather than only fixed categories
- It is open source under AGPL-3.0 with a paid enterprise license for closed-source commercial use — and, as always, real-world accuracy depends on your own data and deployment conditions
