Somebody should check the numbers.

Every open model ships with a benchmark table its own lab wrote. I run the models again — on disclosed hardware, with the harness pinned to a version, and the raw logs pushed to Git. The exotic architectures (MoE, Mamba, hybrids, diffusion LMs) that big leaderboards skip get tested too.

It's all free and reproducible. Sponsorship goes straight to the GPU hours that make the next eval possible. See the work at sanitybench.com.