Why Ultralytics Platform

Link to this sectionThe lifecycle#

Every production CV system goes through the same lifecycle. Ultralytics Platform aims to be the place that handles all of it — see also the steps of a CV project guide for the same lifecycle without the Platform framing:

graph LR
    A[Source data] --> B[Annotate]
    B --> C[Curate / split]
    C --> D[Train]
    D --> E[Validate]
    E --> F{Good enough?}
    F -- no --> A
    F -- yes --> G[Deploy]
    G --> H[Monitor]
    H -- drift --> A

Most teams have a tool (or a script, or a Notion doc) for each step. Each transition between tools is a handoff that loses context — which images were sampled? which labels are stale? which checkpoint is in production? Platform's value is the transitions, not any single step.

Link to this sectionWhat Platform handles well#

• Data ingestion. Upload images, videos (auto-extracted at 1 fps, up to 100 frames), and dataset archives with content-addressable deduplication — see the data management overview.
• Annotation at scale. Manual annotation for all 6 task types plus SAM-powered smart labeling using SAM 2.1 or SAM 3 for boxes and masks.
• Cloud and local training. Cloud GPU runs without provisioning anything, plus local run streaming via the ultralytics package when you train on your own hardware.
• Experiment tracking. Every run is logged automatically; comparison is built in alongside integrations like Comet and MLflow.
• Dedicated endpoints. One-click deployments in 43 global regions with scale-to-zero behavior.
• Monitoring. Live latency, error-rate, request-volume, and log dashboards per endpoint — see monitoring.

Link to this sectionWhat you'll still want to own#

• Custom data pipelines. If your data comes from a proprietary system (DICOM, OEM cameras), you'll integrate it yourself.
• Custom postprocessing. Counting / heatmap / business-logic pipelines (course 3) usually live in your service, not on Platform — many of these patterns ship as Ultralytics Solutions.
• Compliance-locked workloads. If you can't send data outside a specific region, see lesson 9 on regions and the data privacy glossary entry.

The right way to think about Platform is the default home base for the lifecycle — and an explicit decision about which one or two steps you handle yourself. The Platform quickstart is the shortest path from this model to the UI, and the active learning loop it enables is what makes the second iteration cheaper than the first.

Link to this sectionHow it relates to the open-source ultralytics package#

Platform and the OSS package use the same model formats and the same Python API. A model trained on Platform is a normal .pt you can download, run with from ultralytics import YOLO, and export the same way you would locally.

Platform is not a separate framework. It's a managed home for the workflows you'd otherwise glue together.

That means everything you learned in courses 1–3 carries over directly. Platform just removes the infrastructure that surrounds the model.

Link to this sectionPricing model (as of 2026)#

Platform's free tier covers most learning and small projects. Concretely:

• Sign-up credits: $5 (personal email) or $25 (company email) for cloud training.
• Storage: 100 GB on Free, 500 GB on Pro, unlimited on Enterprise.
• Dedicated endpoints: up to 3 on Free, 10 on Pro, unlimited on Enterprise — basic endpoints are free on every plan today.
• Cloud training: usage-based per GPU-hour (RTX PRO 6000 default, A100 / H100 / H200 / B200 / B300 available; B200 / B300 require Pro or Enterprise).
• Failed training runs are not charged.

Indicative cost from the training docs: 1000 images × YOLO26n × 100 epochs ≈ $5 on RTX PRO 6000; 5000 images × YOLO26s × 100 epochs ≈ $23 on H100 SXM. The training dialog shows a real-time estimate before you click Start.