Send us your llama-bench result
Ran an AI model on your own computer? Paste the table llama-bench prints. We read every result by hand, and nothing goes on the site before that.
Why we ask
Our speeds are estimates worked out from each machine's specs. A result measured on a real machine shows where an estimate is right and where it is off. What we publish is labelled self-reported and sits on the same page as our estimate.
How to measure
- Install llama.cpp.llama-bench comes with it.
- Run it on your model file.For example
llama-bench -m your-model.gguf, with the options you normally use. The table shows them. - Copy the table it prints.It measures reading a 512-token prompt (pp512) and writing 128 tokens (tg128). Copy from the first line that starts with | to the build line.
Not on llama.cpp? Email your numbers and how you measured them to alan@localllmrig.com. We publish an emailed result only after you say yes in a reply, with the name or handle you choose.
Your result
We email this result to ourselves through Resend and keep that email for up to 12 months, or for as long as the result is published (it is our record of your consent), then delete it. We never publish your email address. Privacy notice
What happens next
- We read every result by hand. Nothing is published automatically.
- If it holds up, we may add it to that model's page, labelled self-reported, with your name or handle if you gave one.
- If you left an email address, we may write to ask about your setup.