Who serves the model: makers and hosts

The maker is the lab that trained the model. Many makers run an API for their own models. When the weights are public, other companies can serve the model too. These are hosts.

A closed model has one seller, or a few under license. An open model can have dozens. On 12 Nov 2026, 28 hosts listed DeepSeek V4.1 Flash here.

Each host chooses its hardware, the precision it runs the model at, the context window and the output limit. Each sets its own price on top. So the same model name covers a range of offers.

On 12 Nov 2026, the lowest listed input price for GLM 5.3 here was $0.07 per million tokens, and not every host was near it.

Ask five things. What precision is the model served at? What are the window and the output limit? Where does it run, if location matters to you? Is your data kept or used for training? And what happens when you send more requests than your limit allows?

The first two are in the tables here where the host states them. The rest are in its terms and documentation.

Using one host for every model means one key, one bill and one set of limits to learn. Using the lowest-priced host for each model means several of each. The saving has to be worth the bookkeeping.

Runcrate is a host. It serves 124 models through one OpenAI-compatible API, paid from one prepaid balance that also pays for GPUs and storage. Runcrate serves GLM 5.3 at $0.59 per million input tokens.

This is one of a coordinated collection of Runcrate references. Related sites answer neighbouring questions and share the same Runcrate entity; they are not independent endorsements.