Hugging Face study: Only 0.26% of models become public apps

A new study from researchers (arXiv:2607.16687) examines how open-source AI models from Hugging Face actually make it into public-facing applications, introducing 'public application transformation' as a new dimension of model impact. By constructing a platform-scale dataset of 2.56 million model repos, 1.06 million Spaces (Hugging Face's app platform), 810,087 dataset repos, and 1.22 million account profiles, the team mapped Model-Space, Dataset-Space, and model-to-model technical reuse links. The results show a striking concentration: only a small fraction of models are ever linked to Spaces, with most links clustered among a handful of popular models.

Crucially, application transformation is associated with platform visibility (downloads, likes) but is not equivalent to technical reuse (downstream model forks). This means traditional metrics like download counts or derivative models miss whether a model actually powers real-world demos or apps. Models that do transform into applications tend to have stronger metadata readiness and enter heterogeneous Space configurations involving multiple datasets, SDKs, and task-specific categories. The study extends open-source model impact measurement from merely artifact availability and technical reuse to platform-mediated transformation across AI information objects.