Send your measurement

Page updated 2026-10-21

Ran one of the models RigCheck covers on your own hardware, or tried and it did not work? Paste the output here. We check every report against its log before it appears on a page, labeled as a community report.

Did it run? Yes, it ran: here are the numbers No, it failed or ran out of memory

Model

File or variant e.g. 31B UD-Q4_K_XL; for another model, its name too

Hardware GPU(s) with memory, system RAM or Mac memory, operating system

Engine and version e.g. llama.cpp b6200, Ollama 0.12, LM Studio 0.3.30, MLX, ComfyUI

Context length optional

Benchmark output or log llama-bench output (table or JSON), the timing lines of llama-cli / llama-server, or the error

Anything else optional: settings, what you changed, what failed

Name to credit optional; leave empty to stay anonymous

I agree that RigCheck may publish this report under CC BY 4.0, commercial reuse included, credited to the name above or anonymously.

Before sending, remove anything personal from the log: user names in file paths, host names, tokens or keys. We store what you send here, the day and the licence agreement; no email, IP address or cookie (see Privacy). This page loads no Google Analytics or Microsoft Clarity.

What happens next

Checked first: we compare the numbers with the log and the hardware you describe. Reports the log does not support, duplicates and spam are deleted.
Published as a community report: a checked number goes on the model's page next to the exact setup it was measured on (model file, engine, hardware, context), labeled as a community report and never mixed with official numbers or other setups. Failed attempts count too: they tell others where a setup stops working.
Open data: published reports are shared under CC BY 4.0 like the rest of RigCheck's open data, and added to the model's JSON file where that file carries measurements; that is why the form asks for that agreement. Reports without it are not accepted.
How to measure: for llama.cpp, llama-bench -m model.gguf prints the table we read best; add -p 512 -n 128 or your own context with -d. For other engines, paste the timing or error lines.