Ultralytics Platform

Use ultralytics for local YOLO training and inference. Use the generated ultralytics-platform Python SDK for Platform resources, hosted inference, and AI annotation. It handles authentication, typed responses, retries, and errors.

Read the live contract

Before API work, check the generated API reference or GET https://platform.ultralytics.com/openapi.json. Treat the live OpenAPI document as authoritative when examples disagree. These recipes were checked against API and SDK v0.1.62 on 2026-09-24. Check the SDK source for generated method signatures.

uv pip install -U "ultralytics-platform>=0.1.62"
export ULTRALYTICS_API_KEY=ul_... # Settings > API Keys

Platform() reads ULTRALYTICS_API_KEY. The ultralytics package also reads the key saved by yolo login. Never print or commit a key.

Choose the interface

Live training and ul:// URIs

Pass an owner-qualified project to stream a run:

from ultralytics import YOLO

YOLO("yolo26n.pt").train(data="coco8.yaml", epochs=100, project="owner/project", name="run1")

project= is required. Without it, the callback exits before creating a Platform run. Use the owner prefix for a team workspace.

YOLO("ul://owner/project/model").train(data="ul://owner/datasets/dataset", epochs=100)

SDK

Use a context manager and owner/name paths. Keep returned IDs for operations that require them, including image operations, upload assetId, training modelId, and export IDs.

Responses have resource-specific shapes, not a generic envelope. Create calls return id, owner, and the URL name at the top level. Detail calls wrap the resource under its type, such as dataset. A rename changes the URL name, so use the name returned by the update response.

Read references/recipes.md for live-run diagnosis, finished-run upload, dataset upload, hosted inference, Moondream and other AI annotation, and billable jobs.

Invariants

• Confirm the target workspace with client.account.summary() and read the exact resource before a mutation. Team work requires an API key created in that workspace.
• Upload with a signed URL and PUT using the returned headers. Dataset ingest now verifies and completes the upload itself, so upload.complete is optional for datasets. Models still require it.
• Dataset ingest accepts one source: sessionId, sourceUrl, or a connected-storage reference. Set targetSplit when every incoming image must enter one split.
• Top-level model metrics accepts only the contract's named summary metrics. Per-epoch trainResults[].metrics accepts numeric metric names from results.csv.
• Hosted AI annotation uses a stored image ID and dataset class names. It is not a free-form chat or caption API. Single-image predictions are unsaved, while batch annotation writes labels.
• On 429, wait for Retry-After before retrying. Do not invent fixed sleeps.

Cost and destructive actions

Cloud training, model exports, deployments, and batch image processing can spend credits. Confirm the requested scope and cost before an unapproved billable launch. Do not ask again when the user has already authorized it. Check client.billing.usage_summary() for current usage and plan limits. Training returns cost estimates, but not every create response includes a price.

Get approval for deletes outside the user's authorized scope. Project, dataset, and model deletes move resources to 30-day trash. Image deletion and client.lifecycle.delete_trash are permanent.