YOLO Vision 2026:

Model Training#

Ultralytics Platform provides comprehensive tools for training YOLO models, from organizing experiments to running cloud training jobs with real-time metrics streaming.



Watch: Get Started with Ultralytics Platform - Train

Overview#

The Training section helps you:

  • Organize models into projects for easier management
  • Train on cloud GPUs with a single click
  • Monitor real-time metrics during training
  • Compare model performance across experiments
  • Export to 20 deployment formats (see supported formats)

Ultralytics Platform Train Overview

Workflow#

graph LR
    A[📁 Project]:::start --> B[⚙️ Configure]:::proc
    B --> C[🚀 Train]:::proc
    C --> D[📈 Monitor]:::proc
    D --> E[📦 Export]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef proc fill:#2196F3,color:#fff
    classDef out fill:#9C27B0,color:#fff
StageDescription
ProjectCreate a workspace to organize related models
ConfigureSelect dataset, base model, and training parameters
TrainRun on cloud GPUs or your local hardware
MonitorView real-time loss curves and metrics
ExportConvert to 20 deployment formats (details)

Training Options#

Ultralytics Platform supports multiple training approaches:

MethodDescriptionBest For
Cloud TrainingTrain on Ultralytics Cloud GPUsNo local GPU, scalability
Local TrainingTrain locally, stream metrics to the platformExisting hardware, privacy
Colab TrainingUse Google Colab with platform integrationFree GPU access
Automatic GPU Routing

When you select a GPU cheaper than the RTX PRO 6000 and Ultralytics-managed capacity is free, the Platform runs your job on an RTX PRO 6000 while still billing your selected GPU's hourly rate. Runs can finish sooner and cost less than they would have on the selected GPU — the upgrade never adds time or cost.

GPU Options#

Available GPUs for cloud training on Ultralytics Cloud:

GPUGenerationVRAMCost/HourBest For
RTX 2000 AdaAda16 GB$0.24Small datasets, testing
RTX A4500Ampere20 GB$0.25Small-medium datasets
RTX 4000 AdaAda20 GB$0.26Medium datasets
RTX A5000Ampere24 GB$0.27Medium datasets
L4Ada24 GB$0.39Inference optimized
A40Ampere48 GB$0.44Larger batch sizes
RTX 3090Ampere24 GB$0.46General training
RTX A6000Ampere48 GB$0.49Large models
RTX PRO 4000Blackwell24 GB$0.57Budget Blackwell
RTX PRO 4500Blackwell32 GB$0.64Great price/performance
RTX 4090Ada24 GB$0.69Best price/performance
RTX 6000 AdaAda48 GB$0.77Large batch training
L40SAda48 GB$0.86Large batch training
RTX PRO 5000Blackwell48 GB$0.96Large batch training
RTX 5090Blackwell32 GB$0.99Latest consumer generation
L40Ada48 GB$0.99Large models
A100 PCIeAmpere80 GB$1.39Production training
A100 SXMAmpere80 GB$1.49Production training
RTX PRO 6000Blackwell96 GB$2.09Recommended default
H100 PCIeHopper80 GB$2.89High-performance training
H100 NVLHopper94 GB$3.19Maximum performance
H100 SXMHopper80 GB$3.29Fastest training
H200 NVLHopper143 GB$3.39Maximum memory
H200 SXMHopper141 GB$4.39Maximum performance
B200Blackwell180 GB$5.89Large models (Pro+)
B300Blackwell288 GB$7.39Largest models (Pro+)
GPU Tier Access

B200 and B300 GPUs require a Pro or Enterprise plan. All other GPUs are available on all plans including Free.

Signup Credits

New accounts receive signup credits for training. Check Billing for details.

Real-Time Metrics#

During training, view live metrics across three subtabs:

graph LR
    A[Charts]:::start --> B[Loss Curves]:::out
    A --> C[Task Metrics]:::out
    D[Console]:::start --> E[Live Logs]:::out
    D --> F[Error Detection]:::out
    G[System]:::start --> H[GPU, CPU & Memory]:::out
    G --> I[Network & Disk I/O]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef out fill:#9C27B0,color:#fff
SubtabMetrics
ChartsTask metrics (mAP50, mAP50-95, precision, recall for detection), train/val losses, learning rate
ConsoleLive training logs with ANSI color and automatic error detection
SystemGPU utilization, GPU memory and temperature, CPU, RAM, network and disk I/O
Automatic Checkpoints

The best checkpoint (best.pt, the highest-fitness epoch) is uploaded to the Platform periodically while training runs and again when the run ends, so download, export, and deployment always use the best epoch produced so far. Cancelled runs keep the last checkpoint that finished uploading.

Quick Start#

Get started with cloud training in under a minute:

  1. Create a project in the sidebar
  2. Click New Model
  3. Select a model, dataset, and GPU
  4. Click Start Training

FAQ#

  • Training time depends on:

    • Dataset size (number of images)
    • Model size (n, s, m, l, x)
    • Number of epochs
    • GPU type selected

    The current estimator predicts about 6 minutes for 1000 images, YOLO26n, 100 epochs on RTX PRO 6000, and about 2 minutes for 500 images, YOLO26n, 50 epochs on RTX 4090. Actual duration varies; use the live estimate in the training dialog for the selected dataset and configuration. See cost examples.

  • Yes. Concurrent cloud training limits depend on your plan: Free allows 3, Pro allows 10, and Enterprise is unlimited. For additional parallel training, use remote training from multiple machines.

  • If training fails:

    1. The model is marked failed and the compute instance is terminated
    2. The model page shows an error banner with the captured error, a link to the console output, and a Retry action that reopens the training dialog with the same configuration
    3. A run that stops reporting activity for several hours is automatically marked failed and its compute released
    4. If cloud compute had started, elapsed GPU time is charged; failures before compute starts have no GPU usage charge

  • ScenarioRecommended GPU
    Most training jobsRTX PRO 6000
    Large datasets or batch sizesH100 SXM or H200
    Budget-consciousRTX 4090
Contributors
GLglenn-jocher19LAlaodouya1RIRizwanMunawar1SEsergiuwaxmann1

Comments