Visual Large Language Models (VLLMs) integrate visual and linguistic modal information, demonstrating significant advantages in tasks such as image description and visual question answering. In multimodal training, parameter optimization conflicts often result in the degradation of pure text processing capabilities, a challenge referred to as “catastrophic forgetting of pure text tasks”. This poses a critical limitation in high-reliability applications like dialogue systems. Existing approaches typically encounter trade-offs between efficiency and performance, strong task dependency, and high computational costs. This paper proposes an innovative homologous model fusion framework, DPIMerge, which optimizes the weight merging strategy for multimodal models and pure text enhancement models through the Dynamic Parameter Interpolation (DPI) algorithm. The core innovations include: (1) gradient-guided optimization based on the Frobenius norm to constrain the stability of the parameter space; (2) adaptive tuning of dynamic interpolation coefficients to balance the performance of multimodal and pure text tasks. Experimental results show that the performance of the fused model fluctuates minimally in multimodal tasks, while its performance in pure text tasks has been significantly improved. This study provides an efficient and general solution to the modality conflict problem of multimodal models and offers a practical pathway for balancing performance and efficiency for scenarios such as industrial-grade dialogue systems and cross-modal search engines.