What fits, and what measurably runs best

You have eleven GGUF files on disk and one graphics card. Which of them should you actually be running?

That is two questions wearing one coat, and they cost completely different amounts to answer.

The cheap question

What fits is arithmetic. A GGUF header carries the layer count, the KV head count, the head dimensions and the trained context length. Free VRAM is a number the driver will tell you. Put them together and you know how much KV cache a given context costs and whether the weights plus that cache leave room to breathe.

ArcFlare reads only the header — never the tensor data, in windows, so a 30 GB model costs a few megabytes of I/O. Every model, in milliseconds, without loading anything:

The expensive question

What runs best cannot be computed. It has to be measured, which means loading each model and timing it — minutes per model, and it evicts whatever was resident. So it happens only when you ask, and the result is remembered, because a number that took four minutes to measure should not have to be measured again to be read:

What gets measured: time to load, prefill throughput on a 4k prompt the cache has never seen, generation throughput, and five probes.

What the probes are not

Five probes cannot rank models on how well they write code, and pretending otherwise would be the most useful-looking and least true thing this tool could do.

They catch something narrower and still worth catching: a model that has stopped doing the things a harness needs of it — obeying an exact output format, emitting clean JSON, arithmetic, writing a function that actually runs, and finding a fact 4k tokens back.

That is why the recommendation is gated on the probes rather than scored against them. A model that fails them is unusable at any speed, and no amount of tokens per second changes that.

A measurement of a file that has since changed is flagged, not trusted.