Dataset versioning, testable models and an API.
Versioned data for images and video. Import COCO, VOC or YOLO, label the gaps, train four model types, check each version's accuracy and test it on new data, then call the winner over REST. Export when you want your own stack (Pro and Ultra).
Versioned datasets that ship better models.
Reproducible models
Every model points to the exact data version it learned from.
Ship the better version
Accuracy and side-by-side tests before you switch.
Less glue code
Four REST calls instead of a serving stack to maintain.
5 steps, in the real app.
Outputs you can use.
Data lineage
Each model tied to a version
Model accuracy
Per model and version
REST endpoints
Deploy, status, inference, stop
Portable data
Up to 27 export formats
Instead of notebooks, scripts and folders called final_v3…
Pick your role.
Leaving Ultralytics HUB
HUB shut down on July 31, 2026.
- Bring YOLO labels with the images
- Train and evaluate without managing GPUs
- Export YOLOv8 again whenever you want
Moving off Azure Custom Vision
Custom Vision retires on September 25, 2028.
- Bring images with COCO, VOC or YOLO labels
- Accuracy and Test to compare with your old model
- REST calls: run, check, infer, stop
Picking the better version
A new version looked better in training.
- Accuracy for each model and version
- Test both versions on the same new images
- Ask Vero: “Compare v3 and v4”
When it fits, and when it doesn’t.
A good fit
- Baselines, labeling and testing without MLOps glue
- Small teams that want a model behind an API
- Migrating off retired AutoML tools
Use something else when
- You need custom architectures or detailed hyperparameters
- You need on-prem training
- You want an SDK or notebooks in CI (REST only today)
Questions, answered.
Is there an SDK?
Not yet. Every call has ready samples in cURL, .NET, Node and Python in the model's API panel.
Can I control the architecture?
No. You choose the model type, a time budget (Fast, Normal, Extensive) and a size (Mini or Plus). Export the data when you need your own training.
How do I avoid lock-in?
Import COCO, VOC or YOLO; export up to 27 formats (depending on task type); call models over REST.
How do I compare versions?
Each model shows its accuracy and the repository version it learned from. Test both on the same new images, or ask Vero to compare them.
Put your visual data to work.
Free plan, no credit card. Bring your own images or start from a sample.