We have refactored the text model configuration to use gemini-3.7-flash for student-facing content and gemini-3.5-flash-lite for internal judgment tasks. By moving classification and routing to the lite tier, we've achieved a significant reduction in time-to-first-token—dropping from a 2802ms mean to 943ms—without any loss in accuracy. This transition removes dependencies on volatile preview models, ensuring a more stable and performant production environment.