It's unfortunate how those results are presented. The worst result on the pybenchmark link is comparing a numpy based implementation to one that doesn't use numpy.

If the latter can't use numpy or would be slowed down by it, that's a fair comparison, but it looks like they are just at different stages of optimization.