DeepSeek founder Liang Wenfeng has published a new paper with a research team from Peking University, outlining key technical directions for next-generation sparse large language models. The study is widely viewed as an early look at the architecture behind DeepSeek V4.