How does the find-hardware tool work?

You provide a model, quantization, minimum context length and a target tokens-per-second speed. The tool searches measured benchmarks and ranks setups that meet your requirement.

What counts as the 'cheapest' setup?

We use total VRAM as a proxy for cost. The smallest card (or combination) that reaches your target speed wins.

What if no setup reaches my speed?

Try lowering the target, using a smaller quantization such as Q4_K_M, or removing the context-length constraint.

KarimehNet

Search / Log in Register

Find the right hardware

Tell us the model and speed you need. We find the cheapest setup that reaches it.

Start with a model

Pick a model and (optionally) a quantization, context size and minimum speed.