A startup is developing a real-time language translation feature for a low-cost, handheld device. They have developed a large, highly accurate 'teacher' model that achieves state-of-the-art translation quality but requires significant computational resources. They use this model to train a much smaller 'student' model. The student model runs efficiently on the handheld device but occasionally makes minor grammatical errors that the teacher model would not. The company's primary goal is to ensure the device is affordable and has a long battery life, making it accessible to a wide audience. As a consultant, would you recommend deploying the smaller student model or finding a way to use the larger teacher model (e.g., via a cloud API, which would introduce latency and data costs)? Justify your recommendation by evaluating the trade-offs between the two approaches in the context of the company's goals.