A Comparative Analysis of Corporate vs. Individual Participation in Open-Source AI: Evidence from Hugging Face
This paper examines whether corporate organizations and individual contributors behave differently on Hugging Face Hub, the dominant open-source platform for sharing AI models. Using the full population of 518,034 English-language NLP models collected in February 2026, I classify 492,271 models by contributor type via the platform's native user API and estimate a series of OLS and logit regressions with task fixed effects, model age, size, and variant type as controls. The results support a structured division of labour consistent with the private-collective innovation model. Organizational models attract approximately 114% more downloads than individually authored models of comparable characteristics (β = 0.761, p < 0.001), a premium that persists after controlling for dataset documentation. Organizations build larger models (+21.7% in parameter count), use benchmark datasets more systematically, and concentrate in base and pretrained model production (+15.2 percentage points). Individuals dominate fine-tuning (−22.5 pp) and LoRA adaptation (−18.3 pp). Two hypotheses are rejected in unexpected directions: individuals are more likely to use synthetic instruction-tuning data, reflecting the democratization of data generation pipelines, and certain organizations quantize at industrial scale, making quantization a shared rather than purely grassroots activity. The findings extend the private-collective model to trained AI model artefacts and suggest that corporate actors function as platform builders while individuals function as platform extenders, with the two groups operating in a complementary rather than competitive structure.