My understanding is that this can't be done without modifying the source code. If you have evidence to the contrary, I (and others, I'm sure) would like to see.

The current build from Bioconda w/options:

MAFFT v7.407 (2018/Jul/23)
https://mafft.cbrc.jp/alignment/software/
MBE 30:772-780 (2013), NAR 30:3059-3066 (2002)

High speed:
  % mafft in > out
  % mafft --retree 1 in > out (fast)

High accuracy (for <~200 sequences x <~2,000 aa/nt):
  % mafft --maxiterate 1000 --localpair  in > out (% linsi in > out is also ok)
  % mafft --maxiterate 1000 --genafpair  in > out (% einsi in > out)
  % mafft --maxiterate 1000 --globalpair in > out (% ginsi in > out)

If unsure which option to use:
  % mafft --auto in > out

--op # :         Gap opening penalty, default: 1.53
--ep # :         Offset (works like gap extension penalty), default: 0.0
--maxiterate # : Maximum number of iterative refinement, default: 0
--clustalout :   Output: clustal format, default: fasta
--reorder :      Outorder: aligned, default: input order
--quiet :        Do not report progress
--thread # :     Number of threads (if unsure, --thread -1)

L-INS-i is not the recommended method for large alignments:

About

MAFFT is a multiple sequence alignment program for unix-like operating systems. It offers a range of multiple alignment methods, L-INS-i (accurate; for alignment of <∼200 sequences), FFT-NS-2 (fast; for alignment of <∼30,000 sequences), etc.

Also check the Tips for handling a large dataset page.