What is a hosted model?
A hosted model is an AI model that someone else runs. The provider owns the servers, loads the model, keeps it patched and available, and exposes it through an inference endpoint. The customer’s application sends a request over the internet, usually a prompt or a document, and receives the model’s output. Billing is typically token-based pricing or a subscription. The large commercial language models are hosted models, and so are many open-weight models offered by cloud providers and specialist inference companies. The alternative is self-hosted AI, where the company runs the model on infrastructure it controls.
Hosted models in practice
A building products distributor uses a hosted language model to draft product descriptions for its online catalog from supplier spec sheets. The content is public marketing copy, volume is a few hundred items a month, and the team wanted results within a week. A hosted model fits that case well. The same distributor decided against sending customer pricing and order history to the same endpoint and runs that analysis on a self-hosted model instead.
What to watch for
- Read the data terms. Confirm whether prompts and outputs are stored, for how long, in which country, and whether they can be used for training.
- Hosted models change under you. A provider may update or retire a model version, which can alter outputs that workflows depend on.
- Per-token costs scale with volume, so a use case that is cheap during a proof of concept can become a major budget line at production volume.