The MAE is a quantity that is difficult to converge due to different reasons, one of them is that it is tiny. In this context one has to try a lot to gain some experience on how to choose the parameters. I suggest to at least investigate all the parameters mentioned on this site. I also don't know what literature source you take for comparison. Make sure that everything mentioned there is in agreement to what you coose in your calculation, for example the XC functional. In general the FLAPW method implemented in Fleur is one of the most precise implementations of DFT. If you make everything correct you provide the reference in this context. There may, however also be other approaches besides DFT and the FLAPW method implies many parameters so that it is a challenge to make everything correct.

All of the parameters mentioned on the linked site are important. If you have an insulating material the k-point set convergence and the related Fermi smearing are probably not very demanding.

At first I would concentrate on the Kmax convergence, then lmax and lnonsph. Then you could check by adding some local orbitals whether the linearization error is of relevance here and finally I would check out the numbands convergence, which is definitely of relevance for such calculations with spin-orbit coupling.

As an indicator of whether (most) parameters are nicely converged you can also vary the MT radii slightly (reduction by 5-10%) and observe whether this leaves the results stable.