Good small-model tasks
- Classification with a stable label set.
- Extraction into a known schema.
- Short transformations with clear rules.
- High-volume drafts that already have human review.
- Private or local workloads where data cannot leave the environment.
Keep the larger model when
The input is ambiguous, the task requires several kinds of reasoning, the cost of a subtle error is high, or the model must use tools and reconcile conflicting evidence. A cheap model that needs a person to redo the work is not cheap.
How to decide
| Measure | Question |
|---|---|
| Quality | Does the result meet the same approval bar? |
| Correction | How much human work remains? |
| Speed | Does faster response improve the workflow? |
| Cost | What is the cost per approved result? |
| Risk | What happens when the model is wrong? |