i ran a quick experiment where instead of doing boxing the way its done in the beam currently, i used a different boxing (NaN strategy and there was a 10x speedup

Is that translates to real workloads you should open a pr.