Process mining is a domain where computers undoubtedly outperform humans. It is a mathematically complex and computationally demanding problem, and event logs are at too low a level of abstraction to be intelligible in large scale to humans. We demonstrate that if instead the data to mine from are models (not logs), datasets are small (in the order of dozens rather than thousands or millions), and the knowledge to be discovered is complex (reusable model patterns), humans outperform computers. We design, implement, run, and test a crowd-based pattern mining approach and demonstrate its viability compared to automated mining. We specifically mine mashup model patterns (we use them to provide interactive recommendations inside a mashup tool) and explain the analogies with mining business process models. The problem is relevant in that reusable model patterns encode valuable modeling and domain knowledge, such as best practices or organizational conventions, from which modelers can learn and benefit when designing own models.