Any proper benchmarks of Beellama (and its fork Beellama-kvarn) and how it performs quality wise for coding?

So beellama allows you to only reduce quantization on older cache, this should allow better performance than reducing quantization on all cache. They recommend using a 1k tail and even that seems to have big benefits, but I don't see why you couldn't for example use a 20k tail (on a 240k context, th