Deploy computer vision models across 42 global regions
Take trained models from browser testing to production endpoints in a few clicks, with dedicated endpoints, deployment monitoring, and 22 export formats.
Deploy at global production scale
Move trained models into production with worldwide availability, broad export support, and the usage volume proven by the Ultralytics ecosystem.
Test your model in the browser
Every trained model includes a built-in Predict tab functionality. Upload an image, or open your camera; the bounding boxes appear instantly.
Deploy to 42 regions worldwide
Deploy your models to dedicated endpoints across the Americas, Europe, Asia-Pacific, and the Middle East. Each endpoint has its own URL, configurable CPU and memory, and monitoring for requests, latency, errors, and logs.
Endpoints sized for your workload
Start with the included size or choose the CPU and memory each endpoint needs.
- Scale to zero by default: The included size scales to zero when idle and has no uptime charge.
- No Platform API rate limit: Requests go straight to your endpoint, not through the Platform API rate limiter.
- Configurable resources: Choose CPU (1-8 cores) and memory (2-32 GB) to match your workload. Custom sizes keep one warm instance billed for uptime.
22 export formats. Your model. Any environment.
Ultralytics Platform supports cloud and edge deployment for high-performance. All Ultralytics YOLO models are natively optimized to run efficiently across environments, delivering high accuracy, reliable performance, and compatibility even on edge devices with limited compute resources.
Monitor everything in production
Track endpoint requests, latency, errors, and logs from the deployments dashboard. Dedicated endpoints also provide live prediction statistics and temporary examples you can inspect and save to a dataset.
- Requests, latency, and errors: 24-hour HTTP volume, volume-weighted P95 latency, and error-rate alerts above 5%.
- Health checks and logs: Live endpoint health with auto-retry, plus severity-filtered logs to diagnose issues fast.
- Temporary examples: Inspect sampled predictions on dedicated endpoints. Examples are held in endpoint memory; save useful images to a dataset to keep them.
- Prediction statistics: Class counts, confidence, inference time, and spatial heatmaps, where supported by the model task. Statistics reset when the serving instance is replaced.
Integrate in minutes
Every deployed endpoint comes with auto-generated code examples in Python, JavaScript, and cURL, pre-populated with your actual endpoint URL and API key. Copy, paste, and start sending inference requests from any application.
Try YOLO26 Inference
Drag and drop an image to see real-time object detection
Learn how to deploy!
Watch how to test a trained model, deploy it to a global endpoint, and monitor requests, latency, and live predictions.
Need to train a model first?
Ultralytics Platform connects annotation, training, and deployment in a single platform.
Explore industry solutions
See how teams apply Ultralytics computer vision across production environments.
Computer vision in sports
Computer vision for aerial imagery
Computer vision in aerospace
Computer vision in security
Computer vision in robotics
Computer vision in logistics
Computer vision in retail
Computer vision in healthcare
Computer vision in manufacturing
Computer vision in automotive
Computer vision in agriculture
Computer vision in sports
Computer vision for aerial imagery
Computer vision in aerospace
Computer vision in security
Computer vision in robotics
Computer vision in logistics
Computer vision in retail
Computer vision in healthcare
Computer vision in manufacturing
Computer vision in automotive
Computer vision in agriculture
Computer vision in sports
Computer vision for aerial imagery
Computer vision in aerospace
Computer vision in security
Computer vision in robotics
Computer vision in logistics
Computer vision in retail
Computer vision in healthcare
Computer vision in manufacturing
Computer vision in automotive
Computer vision in agriculture
Enterprise-grade security, certified
Independently certified and audited, so your data stays protected in transit and at rest.
SOC 2 Type I
ISO 27001:2022
GDPR & CCPA
Frequently asked questions
Yes. Deploy a model to as many of the 42 regions as your plan allows: 3 endpoints on Free, 10 on Pro and unlimited on Enterprise, counted across your workspace. Deploying closer to your users can reduce latency.
The default endpoint size (1 vCPU, 2 GiB) is free, even while serving requests, and scales to zero when idle. Custom sizes, up to 8 vCPU and 32 GiB, keep one instance warm and are billed from credits for uptime, including idle time, at rates that depend on size and region. Shared inference is included with every plan.
Shared inference runs on a multi-tenant service in 3 regions and is limited to 20 requests per minute per API key or caller, which suits testing. A dedicated endpoint is a single-tenant service in any of 42 regions, with its own URL and API key. Requests sent straight to it aren't subject to that limit, but each endpoint runs one instance and can turn requests away when busy, so load-test before production.
A new endpoint is usually ready in a few minutes, depending on the model and region. Default-size endpoints scale to zero when idle, so the first request after a quiet period can take up to about a minute while the endpoint starts.
Model deployment makes a trained model available to make predictions on new data, such as images or video frames sent from an application. On Ultralytics Platform, you can test a model in the browser, deploy it to a dedicated endpoint in any of 42 regions, call it from your application and monitor it, all from the same workspace.
Send images or supported video files to your endpoint's REST API; depth models accept images only. Each deployment page has Python, JavaScript and cURL examples. Workspace owners can also run an endpoint on a webcam or an internet-reachable RTSP camera from its Predict tab (local network addresses aren't supported), and custom-size endpoints can keep watching one IP camera in the background.
Yes. Export your model for your target hardware and run it on your own servers or edge devices, including offline. Formats include TensorRT for NVIDIA Jetson, OpenVINO, CoreML, LiteRT, NCNN, RKNN, Hailo and Qualcomm; which ones are available depends on the model and task. Using it in proprietary products without AGPL-3.0 obligations requires an Ultralytics Enterprise License.
Start deploying today!
Take your trained models to production across 42 global regions with dedicated endpoints and deployment monitoring.