Recent deep learning-based RIS phase control models achieve competitive performance, but their scalability is constrained by rapidly increasing computational complexity as the number of reflective elements grows. We propose RISnet-LiteMix, a lightweight MLP-Mixer-based RIS phase control model that captures inter-user interactions with significantly lower computational complexity. In particular, the LiteMixer block in RISnet-LiteMix applies dimensional compression or expansion across different perspectives, thereby reducing computational burden while enabling effective learning of channel representations. Consequently, RISnet-LiteMix reduces the number of parameters by 47% and computational complexity by 70% relative to existing MLP-Mixer-based models, without compromising transmission performance.