The advantage of the MSA architecture only truly manifests in long-tail scenarios where context rapidly expands to millions of tokens in long-horizon tasks; it is not particularly fast on short tasks.