Artificial Intelligence (AI) is now ubiquitous. From self-driving vehicles navigating dense city streets to recommendation engines curating personalized content, AI powers applications across nearly every industry. With this broad adoption, assuring the quality of AI models, i.e., the machine-learning components that power these applications, is essential for building reliable and trustworthy systems. Historically, correctness has been the primary focus in AI model development. Yet industry AI models may require many other critical quality attributes. To understand industry perceptions, challenges, and solutions regarding different quality attributes of AI models, we identify nine key quality attributes and conduct interviews with fifteen AI industry practitioners from various countries, companies, and roles. Through the interviews, we summarize practitioners’ perceptions on the importance of quality attributes; for example, efficiency is priorized over correctness for real-time AI applications, and scalability and deployability are no longer AI developers’ primary concerns. Among the challenges we identify, data imbalance stands out as one of the major obstacles to maintaining model correctness and robustness & security; accordingly, our interviews reveal mitigation strategies such as active learning for data collection. We also validate our key findings via a survey of 50 AI industry practitioners, and most findings are well-acknowledged except one being marginally acknowledged. Perceptions of each quality attribute can guide researchers to focus on attributes valued by practitioners when designing new techniques, while avoiding approaches that improve one attribute at the expense of others deemed more critical.