A low-latency option that keeps models in memory for faster single-line translations, though it requires more manual memory management.
