Which AI can tell Chinese stories well? A multi-model evaluation study on the short science fiction generation ability of Chinese large language models

Existing research on the literary generation capabilities of artificial intelligence (AI) has mostly centered on the English-language context. While this has certainly provided an important reference, it may also obscure the differences across language systems. This article takes short Chinese online science fiction as the evaluation object and, based on the perspective of prompt engineering and multi-model comparison, constructs a five-dimensional quantitative evaluation system to systematically test and analyze the story generation abilities of four general Chinese large language models (LLMs): DeepSeek, Doubao, KIMI, and Wenxin Yiyan. By adjusting structured prompts and temperature parameters, combined with a human-AI collaborative scoring mechanism, it evaluates their performance in aspects such as creative setting, character construction, structural coherence, stylistic fluency, and philosophical depth. The results show that DeepSeek exhibits significant advantages. Further cyberpunk rewriting experiment on The Legend of the White Snake shows that DeepSeek can, to a certain extent, recode specific cultural imagery. This article argues that the literary generation capability of AI should not be evaluated merely by plot completeness or linguistic fluency; rather, it should also be examined in terms of whether it can enter the narrative paradigms and cultural expectations of a specific linguistic community. The evaluation of Chinese local LLMs therefore has independent methodological significance.