Train vision AI models in clicks, not days

Yes. Install a current ultralytics package (Python 3.11 or newer), set your Platform API key, and train with project="username/project" to stream metrics to Ultralytics Platform alongside your cloud runs. This works on local machines and in notebooks such as Google Colab. Training on your own hardware uses no Platform credits and doesn't count toward your cloud training limit.

Choose based on model size, dataset size, memory needs and budget. Platform offers 26 NVIDIA GPUs starting at $0.24 per hour. The RTX PRO 6000 is the recommended default for most jobs, and budget GPUs suit smaller experiments. Free workspaces can use 24 of them; B200 and B300 need a Pro or Enterprise plan. The training dialog shows current prices and estimated time and cost before you start.

You pay for cloud GPU time used. A run that fails before its GPU starts costs nothing; once the GPU is running, completed, canceled and failed runs are billed for the elapsed time. If your credit balance drops below zero, active cloud runs stop and are billed for the time used, and auto top-up helps prevent this. A stopped run keeps its latest uploaded checkpoint, if one exists, and Retry Training restores the failed run's model, dataset and training settings.

Yes. Free workspaces can run up to 3 cloud training jobs at once, Pro up to 10, and Enterprise has no limit. Training on your own hardware doesn't count toward these limits.

It depends on dataset size, model size, image size, epochs and GPU. For example, with default settings the estimator predicts about 6 minutes and $0.23 to train YOLO26n on 1,000 images for 100 epochs at 640 pixels on an RTX PRO 6000; larger models take longer. Actual time and cost vary, so check the estimate for your configuration before you start.

Model training optimizes a computer vision model's parameters using labeled examples, so it can recognize the same patterns in new images. On Ultralytics Platform, you go from annotation to training in one workspace: select your dataset, choose a model and a cloud GPU, and start a run.

Yes. Start a run from an official YOLO26, YOLO11, YOLOv8 or YOLOv5 model, from a model you trained earlier, or from an Ultralytics .pt checkpoint you upload (up to 1 GB). To improve a model with new images or classes, add them to your dataset and train from your existing model. This fine-tunes a new model and leaves the original unchanged.

Start with a small Ultralytics YOLO26 model, such as YOLO26n or YOLO26s. Smaller models train and run faster and suit edge devices; larger ones can be more accurate but need more compute. Compare a small and a larger model in the same project, then choose based on accuracy, speed and the device you'll run it on.

Yes. The Free plan has no subscription fee and includes $5 in signup credit, or $25 in total once you verify a company email. Cloud training uses credits for the GPU time used, and training on your own hardware has no Platform charge. See pricing for plans and credits.

Ultralytics YOLO models are licensed under AGPL-3.0 by default, and Free and Pro plans don't include an Enterprise License. To use your models in proprietary products without AGPL-3.0 obligations, you need an Ultralytics Enterprise License. See licensing for details.