Ultralytics

Ultralytics is the open-source team behind the YOLO series of visual models. Officially, it claims to enable one-stop annotation, training and deployment, supports YOLOv5 to YOLO26, covers detection, segmentation and pose estimation, and a commercial license for enterprise use is required for closed-source commercialization.

AI Image Recognition#YOLO#Object Detection#Model Training

Updated: 2026-10-26

What is Ultralytics?

Ultralytics is the maintenance team for the YOLO series of visual models. Its official website positions itself as an "end-to-end computer vision platform", where annotation, training and deployment are completed in one place. Its biggest difference from most online visual services lies in licensing: Ultralytics YOLO is open-source under AGPL-3.0, and the code, models and training processes can be fully run on your own machines and private environments; if used for closed-source commercial products, internal tools, SaaS or embedded devices, you need to purchase an Enterprise License. The official website also provides two paths: a cloud platform and an open-source library. Small teams can first get the process running locally before deciding whether to move to the cloud.

Who is it for?

Algorithm engineers engaged in object detection and industrial vision who need a locally trainable and commercially controllable model stack
For universities and research institutions, the AGPL-3.0 license can be directly adopted for academic or fully open-source scenarios
Manufacturing and quality inspection teams that need to implement production line defect detection and safety compliance monitoring
For robotics, drones, and autonomous driving, which require edge or embedded deployment
Business parties in retail, logistics, agriculture and other sectors that require inventory checking, sorting, and growth monitoring
For beginners who want to learn computer vision from scratch, they can run the examples locally

Key features

Smart Annotation

The official website states that it supports 6 types of annotation tasks including bounding boxes, polygons, semantic masks, key points and rotated boxes. After integrating SAM, it can generate precise masks with one click, supports formats such as YOLO and COCO, and provides team review and version management.

Cloud Training

The official claims 26 cloud GPU configurations are available, ranging from RTX 2000 Ada to NVIDIA B300. It natively supports YOLOv5, YOLOv8, YOLO11 and YOLO26. You can view training metrics in real time and conduct experiment comparisons, with both a no-code interface and a Python SDK provided.

Deployment and Export (Deploy)

The official website states that it supports auto-scaling deployment across 42 global regions and offers 20 export formats including ONNX, TensorRT, CoreML, LiteRT, QNN, Hailo, and Ascend, covering edge, mobile, and embedded hardware.

Multi-task Model Family

The official website lists seven types of tasks including object detection, instance segmentation, semantic segmentation, depth estimation, image classification, pose estimation, and rotated box detection.

Open Source and Enterprise License

AGPL-3.0 applies to academic research, coursework, personal projects and fully open-source projects; the official lists scenarios requiring an enterprise license, including internal tools, commercial products, SaaS, hardware embedding and unopened-source R&D,

Industry Solutions

The official website provides scenarios based on aerial imagery, security, robotics, logistics, retail, healthcare, manufacturing and agriculture. The manufacturing sector lists quality control and defect detection, while the security sector lists perimeter monitoring and license plate recognition.

How to use Ultralytics

First, run the inference of the official pre-trained model locally using open-source libraries to confirm the task type (detection, segmentation, pose estimation, or rotated bounding box), and then prepare the data.
For annotation, it is recommended to use SAM's one-click masking for initial labeling and then manual edge correction, which saves time compared to drawing bounding boxes from scratch.
For training, first run a small number of epochs on a small-scale dataset to verify the workflow, then switch to cloud GPUs for scaling up.
Cloud services are billed by usage, so estimate the cost before formal training.
The most important thing to note is the license: as long as you use Ultralytics' code, architecture, or training outputs, enterprise licensing is required for closed-source commercial use, and the same applies to self-developed models for internal use only.
This should be confirmed at the project initiation stage.
When exporting to edge devices, first test on the actual device using the target format.
prices and terms may be adjusted, so recheck the official website before commercial use.