Ultralytics YOLO27:

Deployment#

Ultralytics Platform provides comprehensive model deployment options for putting your YOLO models into production. Test models with browser-based inference, deploy to dedicated endpoints across 42 global regions, and monitor performance in real-time.



Watch: Get Started with Ultralytics Platform - Deploy

Overview#

The Deployment section helps you:

  • Test models directly in the browser with the Predict tab
  • Deploy to dedicated endpoints in 42 global regions
  • Monitor request metrics, logs, health checks, and temporary predictions on paid endpoints
  • Choose resources: default endpoints scale to zero; custom sizes keep a warm instance with uptime billing

Ultralytics Platform Deploy Page World Map With Overview Cards

Deployment Options#

Ultralytics Platform offers multiple deployment paths:

OptionDescriptionBest For
Predict TabBrowser-based inference with image, webcam, and examplesDevelopment, validation
Shared InferenceMulti-tenant service across 3 data regionsLight usage, testing
Dedicated EndpointsSingle-tenant services across 42 regionsProduction, low latency
ExportDownload weights in 20 formats for local or edge runtimeOffline, on-device

Workflow#

graph LR
    A[✅ Test]:::start --> B[⚙️ Configure]:::proc
    B --> C[🌐 Deploy]:::proc
    C --> D[📊 Monitor]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef proc fill:#2196F3,color:#fff
    classDef out fill:#9C27B0,color:#fff
StageDescription
TestValidate model with the Predict tab
ConfigureSelect a model, region, CPU, and memory; review pricing and the deployment name
DeployCreate a dedicated endpoint from the Deploy tab
MonitorInspect endpoint metrics and temporary predictions in Monitoring

Architecture#

Shared Inference#

The shared inference service runs in 3 key regions. Requests to a model are routed to the service in that model's data region, so results stay inside the region where the model is stored:

graph TB
    User[User Request]:::start --> API[Platform API]:::proc
    API --> Router{Model Data Region}:::decide
    Router -->|US models| US["US Predict Service<br/>Iowa"]:::out
    Router -->|EU models| EU["EU Predict Service<br/>Belgium"]:::out
    Router -->|AP models| AP["AP Predict Service<br/>Taiwan"]:::out

    classDef start fill:#4CAF50,color:#fff
    classDef proc fill:#2196F3,color:#fff
    classDef decide fill:#FF9800,color:#fff
    classDef out fill:#9C27B0,color:#fff
RegionLabelLocationBest For
USAmericasIowa, USAAmericas users, fastest for Americas
EUEurope, Middle East & AfricaBelgium, EuropeEuropean users, GDPR compliance
APAsia PacificTaiwan, Asia-PacificAsia-Pacific users, lowest APAC latency

Dedicated Endpoints#

Deploy to 42 regions worldwide on Ultralytics Cloud:

  • Americas: 14 regions
  • Europe: 13 regions
  • Asia-Pacific: 12 regions
  • Middle East & Africa: 3 regions

Each endpoint is a single-tenant service with:

  • Configurable CPU and memory with pricing shown before deployment
  • Scale-to-zero for default resources; a warm, uptime-billed instance for custom resources
  • Unique endpoint URL with its own interactive API reference at /docs
  • Its own API key binding, so only that key can call the endpoint
  • Independent monitoring, logs, and health checks

Deployments Page#

Access the global deployments page from the sidebar under Deploy. This page shows:

  • World map with deployed region pins; click a region to open the New Deployment dialog
  • Overview cards: HTTP Requests (24h), Active Deployments, HTTP Error Rate (24h), HTTP P95 Latency (24h)
  • Deployments list with three view modes: cards, compact, and table
  • New Deployment button to create endpoints from any completed model
  • Refresh button and an Updated timestamp in the page header

Ultralytics Platform Deploy Page Overview Cards And Deployments List

Automatic Polling

The page refreshes automatically, polling faster while deployments are in a transitional state (creating, deploying, or stopping). See Monitoring for details.

Lightweight, Temporary Monitoring

Paid endpoints provide charts and example images held only in the serving instance's memory. Stopping, restarting, redeploying, resizing, or replacing the model can clear this data. Save examples to a dataset and wait for ingestion to finish to keep them. See Monitoring.

Key Features#

Global Coverage#

Deploy close to your users with 42 regions covering:

  • North America, South America
  • Europe, Middle East, Africa
  • Asia Pacific, Oceania

Scaling Behavior#

Endpoints currently behave as follows:

  • Default resources: idle endpoints scale down to zero and cold-start on the next request
  • Custom resources: one instance stays warm and accrues uptime charges until stopped
  • Single active instance: each endpoint currently serves from one instance on all plans
  • Load shedding: requests receive 429 responses when the endpoint is temporarily at capacity — see Direct Endpoint Requests
  • Request timeout: direct endpoint requests have a maximum duration of 1 hour; see Inference for supported inputs

Regional Deployment#

Use the measured region latency to place an endpoint near its callers. Actual inference latency depends on the model, input size, endpoint state, and network path.

Health Checks#

Each running deployment includes an automatic health check with:

  • Live status indicator (healthy/unhealthy)
  • Response latency display
  • Auto-retry when unhealthy, stopping once healthy
  • Manual refresh button

Quick Start#

Create a deployment:

  1. Train or upload a model to a project
  2. Go to the model's Deploy tab
  3. Click Deploy for a region in the latency table
  4. Review CPU, memory, pricing, and name in the deployment dialog
  5. Click Create Deployment and wait for the deployment status to become Ready
Quick Deploy
Model → Deploy tab → Deploy in a region → Review configuration → Create Deployment

The deployment name is generated from the model name and region city and can be edited before deployment. Once deployed, use the endpoint URL with your API key to send inference requests from any application.

FAQ#

  • FeatureSharedDedicated
    ServiceShared across Platform usersDedicated to one deployment
    ScaleManaged by PlatformDefault: scales to zero; custom: stays warm
    Regions3 data regionsChoose from 42 deployment regions
    URLPlatform model APIGenerated deployment endpoint URL
    TestingModel Predict tabDeployment-card Predict tab or API
    Rate limits20 requests/minuteNo Platform rate limit on direct calls
    AuthAny workspace API keyOnly the API key bound to the deployment

  • The deployment remains in a creating or deploying state while its service starts. It becomes usable when the status changes to Ready; timing varies by model and region, and typically takes a few minutes.

  • Yes, each model can have multiple endpoints in different regions. Deployment counts are limited by plan: Free 3, Pro 10, Enterprise unlimited. The quota is charged to the workspace that owns the model, and an endpoint serves exactly one model at a time — use model replacement to swap it without changing the URL.

  • Default-resource endpoints scale to zero when idle:

    • Endpoint scales down after inactivity
    • First request triggers cold start
    • Subsequent requests are fast

    First requests after an idle period trigger a cold start. Opening the deployment card runs a health check that warms the endpoint, so a test prediction right after it responds quickly.

    Custom-resource endpoints stay warm and are charged for uptime, including idle time. Stop an endpoint to stop its uptime charge.

Contributors
GLglenn-jocher18SEsergiuwaxmann2RAraimbekovm1RIRizwanMunawar1

Comments