AICloudCompareNederlands

WORKLOAD GUIDE

Train or fine-tune a model

The right setup depends on your training method, dataset and memory needs. An inference configuration is not automatically a training configuration.

Find my cloud route →What should I check?

Four checks before you commit.

Start with your requirements. A memory figure alone does not prove a workload will run well.

Training method
Define whether you need full training, full fine-tuning or a parameter-efficient method.
Recoverable work
Save checkpoints and test resuming a job before using interruptible capacity.
Data preparation
Account for cleaning, uploading, access controls and the right to use your training data.
Evaluation
Set a baseline and acceptance criteria so that a completed job is not mistaken for a better model.

Choose how much you want to manage.

These routes explain model serving. Check support for your exact model and requirements.

Training needs its own supported workflow.

The routes below explain who operates a model-serving setup. For training, confirm a separate training or fine-tuning service, or configure a compatible training runtime yourself.

LESS INFRASTRUCTURE TO MANAGE

Hosted model API

Your app sends requests to a provider that already serves the model.

Responsibilities and costs
You manage
Connect the API and handle responses.
Control
Choose from the provider’s supported models and settings.
Billed for
Often tokens, requests or another usage unit.
Check first
Model availability, rate limits and data terms.
Find a suitable route for Hosted model API →

CHOOSE A MODEL AND A SERVICE

Managed endpoint

Deploy a model to a managed serving environment.

Responsibilities and costs
You manage
Configure the model, hardware and scaling.
Control
More deployment choices; access varies by service.
Billed for
Often provisioned compute time; check idle billing.
Check first
Supported runtimes, cold starts and minimum capacity.
Find a suitable route for Managed endpoint →

MORE CONTROL, MORE OPERATIONS

Self-managed runtime

Run your serving software on cloud compute you configure.

Responsibilities and costs
You manage
Deploy, secure, monitor and maintain the runtime.
Control
Choose compatible software and instance settings.
Billed for
Compute, storage, transfer and related services.
Check first
Memory fit, reliability and maintenance effort.
Find a suitable route for Self-managed runtime →
What makes a complete configuration?
01
Model

Name, revision, licence and intended task.

02
Runtime

Serving software, dependencies and deployment support.

03
Compute

Enough usable memory, suitable hardware and capacity.

04
Service

Region, access, storage, monitoring and availability.

Products to investigate.

3 alphabetical catalog examples tagged for this goal. These are starting points to verify, not a ranking or a guarantee of model support.

We currently earn no commission from these links. How we fund this site

Find options for my project →Estimate monthly costsExplain GPU memory →

OFFICIAL DOCUMENTATION / FURTHER READING

AWS: Spot Instances and interruptions (opens in a new tab)Hugging Face: repository licences (opens in a new tab)Hugging Face: endpoint billing (opens in a new tab)

Editorial references checked 11 October 2026. Product features and terms can change; confirm the current documentation before purchasing.