Run many independent subtasks in parallel on cheap, local or open-source models and return only the distilled result.
