This one ends up somewhere in the middle:

sub tux_for {
    my ($length, $ones) = @_;
    [   map  { substr unpack ("b*", $_), 0, $length }
        grep { $ones == unpack "%32b*" => $_ }
        map  { pack "L<", $_ }
        0 .. oct "0b".join""=> (1) x $ones, (0) x ($length - $ones)
        ];
    }
[download]

This one is 1200 x faster than Discipulus, but still 1200 x slower than the rest:

use Algorithm::FastPermute;

sub tux_a_fp {
    my ($length, $ones) = @_;
    my @l = ((0) x ($length - $ones), (1) x $ones);
    my %seen;
    permute { $seen{join "" => @l}++ } @l;
    [ keys %seen ];
    }
[download]