Which LLMs are best at low-latency translation?