Precision and the price of a model

A model is a very large set of numbers. Stored at full precision, each number takes sixteen bits or more. A host can store them in eight bits or four instead. The model then needs less GPU memory and runs faster, which lowers the host’s cost.

This is called quantization. The listings this site reads label it with short codes such as bf16, fp8 and fp4. The number is the count of bits. A smaller number means a more compressed model.

A host that fits the model in less memory serves more requests from the same GPU. It can charge less per token. So the lowest rates for a model are often reduced-precision listings.

In the listings read for this site, Qwen3.8 27B appears at fp4, fp8, bf16 and fp16. Those are four ways of serving one model, and their rates are not the same.

A compressed model is not identical to the original. For many tasks the difference is small. For others, such as long reasoning chains, exact code or tool calls with strict formats, it can show.

No price page tells you which case is yours. The only test is your own: run a sample of real prompts against the listing you plan to buy and read the answers.

Compare a rate only with rates at the same precision. Put an fp4 price beside an fp8 price and the lower number looks like a saving, when it is a different service.

Some listings carry no precision label. Treat those as unknown, not as full precision. Ask the host, or test.

Runcrate puts 125 models behind one OpenAI-compatible API, paid from one prepaid balance. That makes the test cheap to run: point your existing client at it, send the prompts you care about and judge the output yourself. Runcrate’s live rate for each model in the tables on this site is read from its own feed.

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.