This is a submission for the Gemma 4 Challenge: Write About Gemma 4 I built a local vision project with Gemma 4 where a small model runs on an edge device and a bigger model runs on a stronger local machine. The small model is fast and private. The bigger model is slower, but better at careful reasoning. That setup taught me something useful: Fine-tuning should not be the first thing you reach for. Before collecting a dataset, launching a training job, or changing weights, try this: Use a larger Gemma 4 model as a teacher to improve how you prompt and route a smaller Gemma 4 model. This post walks through the pattern I used: prompt upskilling, escalation, and knowing when fine-tuning is actually worth it. The Problem: Small Models Are Fast, But Sometimes Too Confident Small local models are exciting because they make edge AI feel practical. You can run inference close to the sensor, avoid sending every input over the network, and keep latency low.…