Strictly speaking, MoarVM is not an NQP interpreter - it runs MoarVM bytecode.

Sure, MoarVM is by design semantically close to NQP, but if you look at

    $ nqp-m --stagestats -e ''
    Stage start      :   0.000
    Stage parse      :   0.003
    Stage ast        :   0.000
    Stage optimize   :   0.000
    Stage mast       :   0.005
    Stage mbc        :   0.000
    Stage moar       :   0.000

you get mostly the same steps as on the JVM

    $ nqp-j --stagestats -e ''
    Stage start      :   0.000
    Stage classname  :   0.045
    Stage parse      :   0.067
    Stage ast        :   0.000
    Stage optimize   :   0.009
    Stage jast       :   0.145
    Stage classfile  :   0.010
    Stage jar        :   0.000
    Stage jvm        :   0.003

except that MoarVM bytecode files do not get bundled into JARs and there's no need to do the extra processing that happens in the 'classname' stage.