The next generation of large language models focuses on optimization for excellent reasoning, multi-task knowledge, and multilingual.