# Hosted Fine-Tuning versus Open-Source Training — Fine-tuning vs Prompting

Source: https://www.skillbyai.com/en/fine-tuning/pr-platform

> Choose by control, data sensitivity, effort and long-term cost.

## Rent the cooking or own the kitchen

**Hosted fine-tuning** from a model provider means you upload data and get a model id: least operational work, but you accept the provider's supported models, settings, pricing, data terms and retirement schedule, and the weights are not yours. **Open-weight models** trained with open-source tools on your or rented GPUs give full control over data location, weights, adapters and serving, at the price of ML and infrastructure work. Pick on data sensitivity and residency, how much customisation you need, team skills, expected volume and how long you must keep a model running. Supported models and features change quickly, so read the current documentation.

## Hosted versus open-weight

Typical trade-offs; confirm with current terms.

```text
                       hosted fine-tune              open-weight + own training
effort                 low                           higher (ML + infra)
control of weights     provider holds them           you hold them
data location          provider terms apply          you decide
model choice           provider's supported list      many open models
long-term risk         model retirement, price change hardware, upkeep, security
best for               fast start, moderate volume   sensitive data, high volume, deep customisation
```

## Read the data terms

Check whether uploaded training data is retained or used to improve provider models before you upload anything sensitive.

**Quiz:** What is a key advantage of open-weight training?

- [ ] It is always cheaper at any volume
- [ ] It needs no engineering
- [x] You control the weights, data location and serving
- [ ] It removes the need for evaluation

*Answer:* You control the weights, data location and serving. Control comes with operational responsibility.