Hugging Face User Perspective: A Study on Real-World Usage Experiences of General-Purpose and Multimodal Large Models

An empirical study based on 662 discussion threads from the Hugging Face platform focuses on the real-world usage experiences of users of general-purpose and multimodal large models, revealing core pain points such as access barriers, generation quality, and deployment complexity. By analyzing diverse user feedback, the study provides key references for improving the large model ecosystem.

Large language models have evolved to multimodal, but existing studies have limitations: questionnaires restrict users' free expression and easily miss unforeseen issues; analysis of Reddit/GitHub Issues tends to focus on failure debugging, making it difficult to fully capture the diverse experiences in normal usage scenarios, leading to blind spots in understanding user needs.

Hugging Face was chosen as the research platform because it is a globally important model hosting and collaboration community, bringing together diverse models from academia and industry as well as active discussions. The study collected 662 discussion threads for 38 representative models (21 general-purpose, 17 multimodal), and constructed a three-level taxonomy through manual annotation to systematically categorize user concerns.

Access barriers are one of the most prominent issues for users: difficulties in model downloading (tens of GB of weights require high-speed networks), API usage restrictions, regional access limitations; unstable networks in resource-constrained regions and unclear license terms for some models limit the inclusivity of the technology.

Generation quality issues include inconsistent outputs, hallucinations, and insufficient understanding of domain-specific knowledge; multimodal models additionally face challenges such as insufficient sensitivity to details in image understanding and mismatches between text descriptions and visual content, which are particularly prominent in high-precision scenarios.

Deployment and invocation complexity hinders widespread application: moving from experiments to production requires solving engineering problems such as dependency management, performance optimization, and service-oriented deployment; multimodal models need to handle preprocessing/postprocessing of different modalities, interaction coordination of subsystems, and higher computational resource requirements, which deters potential users.

Improvement suggestions: At the access level, provide clear terms, segmented downloads/incremental updates; at the quality level, strengthen domain-specific evaluation and optimization; at the deployment level, develop user-friendly tools and standardized interfaces; at the documentation level, improve tutorials, examples, and troubleshooting guides; at the community level, establish a more comprehensive user support mechanism.

This study, based on empirical analysis from the Hugging Face platform, provides valuable insights for understanding the real experiences of large model users and reveals deficiencies at both the technical and ecosystem levels. Addressing these pain points will be key to promoting the popularization and deepening of large model technology applications.