"Muscle" mentions the following:
"2.3 Large alignments If you have a large number of sequences (a few thousand), or they are very long, then the default settings of may be too slow for practical use. A good compromise between speed and accuracy is to run just the first two iterations of the algorithm. On average, this gives accuracy equal to T-Coffee and speeds much faster than CLUSTALW. This is done by the option –maxiters 2, as in the following example.
muscle -in seqs.fa -out seqs.afa -maxiters 2
With varying degrees of accuracy/quality.
I've also done up to about 30kb with CLUSTALO in the past which seemed to work reasonably well.
Kalign and LAST are specifically intended for long sequences though, so start there.