Abstract: Complex distributed hydrological models based on physical factors involve high computational costs, and it is difficult to apply conventional global optimization algorithms to these models because of the large number of model simulations required. Efficiently finding optimal parameters with fewer model executions is therefore crucial for the iterative optimization of complex models. This paper proposes an intelligent interactive parameter optimization framework based on large language models (LLMs) and systematically evaluates the performance of six mainstream LLMs in hydrological model parameter optimization, using the Hydrologiska Byråns Vattenbalansavdelning (HBV) and Variable Infiltration Capacity (VIC) models as case studies. The results show that: ① LLMs, with their deep understanding of the physical meanings of parameters and feedback indicators, achieve 95% of the optimal solution with an average of only 45 iterations, significantly outperforming conventional algorithms, which typically require over 100 iterations; ②LLMs perform excellently in low- to medium-dimensional parameter spaces (≤6 parameters), while their optimization performance declines significantly in high-dimensional parameter tasks, though inference-based models exhibit stronger robustness; and ③ under expert knowledge-guided strategies, the average Nash efficiency coefficient of the VIC model improves by 0.14 compared to zero-knowledge strategies, and the context memory mechanism effectively enhances optimization stability. This study introduces LLMs into the hydrological model parameter optimization process, demonstrating the effectiveness of a diagnosis—feedback—adjustment approach, and provides a reference for paradigm innovation in empowering scientific research with LLMs.