Because "the model" is only the weights. What you get back is the weights plus a serving stack: the precision the weights are stored and computed at, the inference engine and its kernels, how requests are batched, which sampling defaults apply, the context window the provider actually allows, and anything the provider changes after the listing goes up. Two providers can list the same model name and differ on every one of those. Nothing in the response says so.
This page is about open-weight models served through routers and marketplaces, where one model name can sit above a dozen endpoints at prices several times apart. It names no provider. For how Bitdepth measures the difference, see the method.
Weights are released at one number format and can be served at another. bf16 stores 16 bits per value; fp8 stores 8; fp4 and int4 store 4. Fewer bits means less memory, more throughput and a lower price, and in principle less fidelity to the original. Some models are released by their makers at 8 or 4 bits already, so a lower declared precision is not automatically a loss, and a careless 8-bit conversion can lose more than a careful 4-bit one. The declared precision is a format the provider reports, not a measurement of the output. More on what the precision field tells you.
Two endpoints at the same precision can still return different text. The inference engine, the attention kernels, how many requests share a batch, the key-value cache precision, the default sampling parameters when a request leaves them unset, and the maximum context the provider honors all change what comes out. Temperature zero does not make this go away: floating-point sums come out differently depending on batch composition and kernel order, so the most likely token can flip on a close call, and a flipped token early in an answer changes everything after it.
An endpoint is not fixed. Providers change engines, re-quantize, move hardware, and route overflow elsewhere during peak load. The model name stays the same. A buyer who tested an endpoint in one month is holding a result about that month.
Give every endpoint of one model the same short task at temperature zero and read the answers side by side. On a word-reversal prompt, endpoints under one model name split into camps: most return the right letters with a stray space inside, some return nothing at all, a few return the letters in a different order. On an arithmetic word problem that asks for the integer only, most endpoints return the integer, a few return a paragraph of working, and occasionally one returns a different number. The family of task that separates endpoints most is exact string manipulation, which is where precision loss shows first. Arithmetic, structured output and long-context retrieval agree far more often, until one endpoint stops agreeing.
None of this says why. It says the outputs differ, by how much, and whether the difference is larger than the endpoint's disagreement with itself from one night to the next.
Bitdepth does the last two every night, for every endpoint of each model it covers, with the same prompt set, scored against a reference built from the endpoints that declare full precision, and publishes the grade beside the price. How a grade is made.