ultralytics platform
Train Ultralytics YOLO models on 22 cloud GPUs, monitor every metric in real time, and compare experiments side by side, all from a single platform.
127.7K
GitHub start
234M
Downloads
2.5B
Daily usages
Ultralytics Platform is purpose-built for the models you already use. Train Ultralytics YOLO26, YOLO11, YOLOv8, and YOLOv5 across all five computer vision tasks, with full support from nano to large.
Start from an Ultralytics YOLO model: YOLO26, YOLO11, YOLOv8, or YOLOv5 models, pre-trained by the original authors, ready to fine-tune.
Bring your own computer vision model: Upload a .pt file and train it on cloud GPUs. Training arguments, architecture, and results are parsed automatically.
Your dataset or ours: Connect your training data and labelled datasets or browse official Ultralytics and community-shared datasets to get started.
Choose from 22 GPU options, from the RTX 4090 and A100 to the H100, H200, and B200. Select a GPU, set your budget, and start training. The platform estimates cost and duration upfront, so there are no surprises.
Prefer your own hardware? Train on your local GPUs or CPUs and stream real-time metrics back to the platform using the Ultralytics Python package. Your experiments appear alongside cloud runs in the same project dashboard.
Loss curves, mAP, precision, and recall are plotted per epoch, with automatic checkpoints and best model preservation throughout.
Live training logs are streamed from the GPU with ANSI colour support and automatic error detection, so problems surface immediately.
Real-time GPU usage, memory, temperature, CPU, and disk telemetry confirm your GPU is running efficiently throughout the run.
Model validation is a key step once your computer vision models have finished training. Review your confusion matrix, PR curve, and per-class metrics directly in Platform, then export to 17+ formats, optimised for cloud, edge, or on-device deployment.
1
Annotate
2
Train
3
Deploy
Yes. Ultralytics Platform supports local training on your own GPUs or CPUs. Install the Ultralytics Python package, set your API key, and start training, real-time metrics stream directly to the platform dashboard alongside your cloud training runs. This gives you the flexibility to use your own hardware while keeping all experiments organized in one place.
Ultralytics Platform offers 22 GPU options ranging from $0.24 to $4.99 per hour. For most workloads, the RTX PRO 6000 (96 GB, $1.89/hr) is a strong default. For time–sensitive training, the H100 and H200 deliver maximum performance. For testing and small datasets, budget options like the RTX 2000 Ada ($0.24/hr) work well. The platform shows an estimated cost and duration before you start, so you can choose the right balance of speed and budget for your project.
If a training run fails, you won't be charged. You're only billed for actual GPU time on completed or manually cancelled runs. Checkpoints are saved throughout training, so if a run is interrupted or cancelled, your progress up to that point is preserved. You can review console logs to diagnose issues and restart training with adjusted settings.
Yes. Ultralytics Platform supports concurrent training runs. Free plan users can run up to 3 simultaneous training jobs, while Pro plan users can run up to 10 and Enterprise unlimited. Each run gets its own dedicated GPU instance.
Training time depends on your dataset size, model size, number of epochs, and GPU selection. As a reference, training YOLO26n on 1,000 images for 100 epochs takes approximately 2-3 hours on an RTX PRO 6000. Larger models like YOLO26x will take longer for the same configuration. The platform estimates cost and duration before training starts, so you always know what to expect.
Model training is the process of teaching a computer vision model to recognize patterns in visual data. During training, the model processes thousands of labeled images, adjusts its parameters, and progressively improves its ability to detect, segment, or classify objects. On Ultralytics Platform, training is integrated directly into the annotation and deployment workflow. Once your dataset is labeled, you can select a YOLO model, choose a cloud GPU, and start training, all without leaving the platform.
Build production-ready vision AI models on cloud GPUs — starting at $0.24 per hour.
