The largest models which generate the headline benchmarks are never released after any number of years, it seems.

Very difficult to replicate results.