Glenn Jocher d71f7f3266
Platform docs (#23338)
Signed-off-by: Glenn Jocher <glenn.jocher@ultralytics.com>
Co-authored-by: UltralyticsAssistant <web@ultralytics.com>
2026-01-20 11:51:00 +00:00

7.7 KiB

comments description keywords
true Learn how to manage, analyze, and export trained models in Ultralytics Platform with support for 17 deployment formats. Ultralytics Platform, models, model management, export, ONNX, TensorRT, CoreML, YOLO

Models

Ultralytics Platform provides comprehensive model management for training, analyzing, and deploying YOLO models. Upload pretrained models or train new ones directly on the Platform.

Upload Model

Upload existing model weights to the Platform:

  1. Navigate to your project
  2. Click Upload Model
  3. Select your .pt file
  4. Add name and description
  5. Click Upload

Supported model formats:

Format Extension Description
PyTorch .pt Native Ultralytics format

After upload, the Platform parses model metadata:

  • Task type (detect, segment, pose, OBB, classify)
  • Architecture (YOLO26n, YOLO26s, etc.)
  • Class names and count
  • Input size and parameters

Train Model

Train a new model directly on the Platform:

  1. Navigate to your project
  2. Click Train Model
  3. Select dataset
  4. Choose base model
  5. Configure training parameters
  6. Start training

See Cloud Training for detailed instructions.

Model Overview

Each model page displays:

Section Content
Overview Model metadata, task type, architecture
Metrics Training loss and performance charts
Plots Confusion matrix, PR curves, F1 curves
Test Interactive inference testing
Deploy Endpoint creation and management
Export Format conversion and download

Training Metrics

View real-time and historical training metrics:

Loss Curves

Loss Description
Box Bounding box regression loss
Class Classification loss
DFL Distribution Focal Loss

Performance Metrics

Metric Description
mAP50 Mean Average Precision at IoU 0.50
mAP50-95 Mean Average Precision at IoU 0.50-0.95
Precision Ratio of correct positive predictions
Recall Ratio of actual positives identified

Validation Plots

After training completes, view detailed validation analysis:

Confusion Matrix

Interactive heatmap showing prediction accuracy per class:

PR/F1 Curves

Performance curves at different confidence thresholds:

Curve Description
Precision-Recall Trade-off between precision and recall
F1-Confidence F1 score at different confidence levels
Precision-Confidence Precision at different confidence levels
Recall-Confidence Recall at different confidence levels

Export Model

Export your model to 17 deployment formats:

  1. Navigate to the Export tab
  2. Select target format
  3. Click Export
  4. Download when complete

Supported Formats (17 total)

# Format File Extension Use Case
1 ONNX .onnx Cross-platform, web, most runtimes
2 TorchScript .torchscript PyTorch deployment without Python
3 OpenVINO .xml, .bin Intel CPUs, GPUs, VPUs
4 TensorRT .engine NVIDIA GPUs (fastest inference)
5 CoreML .mlpackage Apple iOS, macOS, watchOS
6 TF Lite .tflite Mobile (Android, iOS), edge
7 TF SavedModel saved_model/ TensorFlow Serving
8 TF GraphDef .pb TensorFlow 1.x
9 TF Edge TPU .tflite Google Coral devices
10 TF.js .json, .bin Browser inference
11 PaddlePaddle .pdmodel Baidu PaddlePaddle
12 NCNN .param, .bin Mobile (Android/iOS), optimized
13 MNN .mnn Alibaba mobile runtime
14 RKNN .rknn Rockchip NPUs
15 IMX500 .imx Sony IMX500 sensor
16 Axelera .axelera Axelera AI accelerators

Format Selection Guide

For NVIDIA GPUs: Use TensorRT for maximum speed

For Intel Hardware: Use OpenVINO for Intel CPUs, GPUs, and VPUs

For Apple Devices: Use CoreML for iOS, macOS, Apple Silicon

For Android: Use TF Lite or NCNN for best performance

For Web Browsers: Use TF.js or ONNX (with ONNX Runtime Web)

For Edge Devices: Use TF Edge TPU for Coral, RKNN for Rockchip

For General Compatibility: Use ONNX — works with most inference runtimes

!!! tip "Export Time"

Export time varies by format. TensorRT exports may take several minutes due to engine optimization.

Dataset Linking

Models can be linked to their source dataset:

  • View which dataset was used for training
  • Access dataset from model page
  • Track data lineage

When training with Platform datasets using the ul:// URI format, linking is automatic.

Visibility Settings

Control who can see your model:

Setting Description
Private Only you can access
Public Anyone can view on Explore page

To change visibility:

  1. Open model actions menu
  2. Click Edit
  3. Toggle visibility
  4. Click Save

Delete Model

Remove a model you no longer need:

  1. Open model actions menu
  2. Click Delete
  3. Confirm deletion

!!! note "Trash and Restore"

Deleted models go to Trash for 30 days. Restore from Settings > Trash.

FAQ

What model architectures are supported?

Ultralytics Platform supports all YOLO architectures:

  • YOLO26: n, s, m, l, x variants (recommended)
  • YOLO11: n, s, m, l, x variants
  • YOLOv10: Legacy support
  • YOLOv8: Legacy support
  • YOLOv5: Legacy support

Can I download my trained model?

Yes, download your model weights from the model page:

  1. Click the download icon
  2. Select format (original .pt or exported)
  3. Download starts automatically

How do I compare models across projects?

Currently, model comparison is within projects. To compare across projects:

  1. Transfer models to a single project, or
  2. Export metrics and compare externally

What's the maximum model size?

There's no strict limit, but very large models (>2GB) may have longer upload and processing times.

Can I fine-tune pretrained models?

Yes! Upload a pretrained model, then start training from that checkpoint with your dataset. The Platform automatically uses the uploaded model as the starting point.