Contact

hello@localllmchecker.com

Reporting a number that looks wrong

This is the most useful thing you can send. Our figures are calculated, not measured, so a systematic gap between our estimate and your machine usually means a term in the model needs fixing — and fixing it improves every page at once.

Include, if you can:

• The model and the exact quantisation file you loaded
• Your GPU or machine, and how much memory it has
• The runtime and version, and the context length you set
• What the runtime reported for memory used, and the tokens per second you saw
• What we said instead

Missing hardware

Tell us the device and, ideally, its memory capacity and peak memory bandwidth from the vendor datasheet. Those two numbers are all the calculator needs.

Missing models

Send the Hugging Face repository name. Anything with a published configuration and a conventional attention cache can be added; state-space and hybrid architectures are deliberately excluded, for the reasons set out in the methodology.

Everything else

Press, partnerships and advertising enquiries go to the same address. We do not accept paid placement in the rankings or sponsored entries in the catalogue, so please do not ask.