Is the same model the same on every host?

A model’s weights are numbers. A host can store them at full precision or in a smaller format that needs less memory and runs on less hardware. This is called quantization. It lowers the host’s cost, and it can lower the quality of the answers.

In the listings read for this site, the hosts of a single model state several different precisions. Some state none.

Often the difference is small. Sometimes it shows on exactly the hard cases you care about: long reasoning, careful code, or strict output formats. No general score can settle it for your work.

So treat a lower-precision listing as a different offer. It may be the right one. Decide with your own prompts.

The context window and the output limit are the host’s choice. So are the default settings applied when you send none. So is whether features such as tool calls and image input are switched on.

Two hosts can also serve different dated versions under a similar name. Read the full model name in each row.

Keep a small set of your own prompts with answers you have judged good. Send them to each host of the model at the same settings. Read the answers yourself. Note the token counts each host reports, since those drive the bill.

This site states no quality scores or speeds. It prints what each host publishes and the date it was read.

Runcrate serves 124 models behind one OpenAI-compatible API, so one test script covers them all. One prepaid balance pays for the tests.

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.