In mobile deep learning tasks, particularly for deploying multimodal large language models (mLLMs), the incompatibility between hardware operator support and model complexity poses a key challenge. To address operator parameter constraints (e.g., convolution/pooling kernel size limitations) and operator absence issues in NPU-accelerated multimodal model inference, we propose a dual optimization strategy: 1) An operator composition substitution method based on equivalent computational graphs, achieving NPU-compatible transformation of restricted operators through tensor remapping and weight migration; 2) MLP surrogate networks for unsupported operators, leveraging multilayer perceptrons’ universal approximation capability to reconstruct irregular computational paths. By establishing an operator substitution cost model that comprehensively evaluates the Pareto frontier of computational complexity versus accuracy degradation, we ultimately select the optimal deployment scheme. Experimental results demonstrate that the optimized deep learning model exhibits only 0.5% accuracy degradation on NPU platforms, while significantly reducing inference energy consumption by 85.7–91.6%, 83.3–96.6%, and 95–97.5% compared to CPU, GPU, and NPU-CPU heterogeneous architectures, respectively, validating the method’s effectiveness in balancing computational efficiency and model performance.