【目的/意义】 鉴于数字技术的演进催生知识形态从固态向液态的认知相变，针对现有研究在系统揭示DeepSeek等跨模态推理技术重构知识生产范式机制方面存在的分析框架缺位问题，提出知识液态化理论框架，揭示跨模态推理技术对知识生产范式的重构机制。 【方法/过程】 构建包含“模态—主体—流程”3个维度的知识液态化分析框架，采用案例分析与功能解析法，系统剖析DeepSeek的动态稀疏注意力机制、混合专家（MoE）模型及跨模态对齐技术重构知识生产范式的核心动因。 【结果/结论】 DeepSeek并非一般性的影响因素，而是通过其技术架构（MoE等）与开源模式，从根本上触发知识生产在模态（从离散到统一）、主体（从中心化到人机分布式）、流程（从线性到动态交互）的三重范式重构。这一重构过程在提升知识生产效率的同时也衍生出认知权威耗散、责任归属模糊等新型治理挑战。

[Purpose/Significance] Given that the evolution of digital technology has triggered a cognitive phase transition of knowledge forms from solid to liquid, this study addresses the gap in existing research, where there is a lack of analytical frameworks for systematically uncovering the mechanisms through which cross-modal reasoning technologies such as DeepSeek reconstruct the paradigm of knowledge production. This study proposes a theoretical framework of knowledge liquidization and reveals the reconstruction mechanism of cross-modal reasoning technologies on the knowledge production paradigm. [Method/Process] An analytical framework of knowledge liquidization encompassing three dimensions, “modality-subject-process”, was constructed. By integrating case analysis and functional analysis methods, this study systematically examined the core drivers through which DeepSeek’s dynamic sparse attention mechanism, mixture of experts (MoE) model, and cross-modal alignment technology reconstructed the knowledge production paradigm. [Result/Conclusion] The findings indicate that DeepSeek is not merely a general influencing factor. Instead, through its technical architecture (such as the MoE model) and open-source model, it fundamentally triggers a threefold paradigmatic reconstruction of knowledge production: in the dimension of modality, transforming from discrete forms to unified forms; in the dimension of subject, shifting from centralized dominance to human-machine distributed collaboration; in the dimension of process, evolving from linear progression to dynamic interaction. While this reconstruction process significantly improves the efficiency of knowledge production, it also gives rise to new governance challenges, such as the dissipation of cognitive authority and ambiguity in responsibility attribution.