Deploy a Model

What you get

A Platform deployment is:

• An HTTPS endpoint URL (https://api.platform.ultralytics.com/v1/predict/<deployment-id>).
• An API key for auth.
• Auto-scaled compute behind it (the deployment scales to traffic).
• Built-in latency and accuracy dashboards.
• Versioned: each deployment is tied to a specific run + checkpoint.

Deploy

In the UI: Deploy → New Deployment. Pick:

Click Deploy. After ~30 seconds you have an endpoint.

Calling the endpoint

curl -X POST "https://api.platform.ultralytics.com/v1/predict/<id>" \
  -H "Authorization: Bearer $ULTRALYTICS_API_KEY" \
  -F "image=@test.jpg"

That returns JSON with detections:

{
  "deployment": "forklift_v3",
  "model_version": "v3_pallets_added",
  "detections": [
    {
      "class": "forklift",
      "class_id": 0,
      "confidence": 0.92,
      "box": [120, 234, 480, 700]
    }
  ],
  "inference_ms": 18,
  "request_id": "req_abc123"
}

from ultralytics_platform import Client

client = Client(api_key=os.environ["ULTRALYTICS_API_KEY"])
result = client.predict("forklift_v3", image="test.jpg")
for det in result.detections:
    print(f"{det.class_name} {det.confidence:.2f}")

Latency vs cost

Two knobs you'll touch most:

• Min replicas. Min = 0 is cheapest but gives 5–20 s cold inference-latency starts on first request. Min = 1 keeps a hot replica.
• Hardware tier. GPU vs CPU; smaller GPUs vs bigger. The throughput-vs-latency guide is a good companion when you need to choose between batched and per-request modes.

For real-time clients (browsers, IoT) — including edge AI deployments calling back into Platform — min ≥ 1 is mandatory. For batch jobs that tolerate cold starts, min = 0 is correct and saves a lot of money. The model deployment options and practices guides go deeper on the cost/latency tradeoffs.

Versioned deploys and rollbacks

Each deployment is tied to a specific run. To deploy a new model:

1. Train v4. Validate. Decide to ship.
2. Click "Update deployment" → pick the v4 run.
3. Platform rolls v4 in alongside v3, then switches traffic.
4. If v4 misbehaves, click "Rollback" — instant return to v3.

This is the same gradual rollout pattern Vercel uses for code. Use it; never edit a live deployment by replacing files in place. If you'd rather self-host, Triton Inference Server is the canonical option for a YOLO-based gateway running in your own cloud.

Custom domain

Production usually wants detect.your-company.com instead of the platform-hosted URL. Add a CNAME to the deployment URL and Platform handles TLS.

Billing and quotas

Platform charges per inference (or per replica-hour for always-on deployments). Set quotas:

• Hard quota → endpoint returns 429 above the quota.
• Soft quota → endpoint warns; you can decide policy.

Quotas prevent a runaway client from blowing your budget.