Why open a generative AI model? A typology based on what is open and what is not.

The rapid expansion of generative AI has intensified interest in open models. However “open source AI” is still used to describe systems with very
different degrees of transparency and reuse. Using the Open Source AI Index database, we analyzed 189 models evaluated with Liesenfeld et al.’s openness grid, covering availability, documentation, and access methods through fourteen criteria. We apply Hierarchical Clustering on Principal Components (HCPC) in R (FactoMineR) to identify openness profiles. The analysis yields five clusters: “open washing”, “easy access”, “open weight”, “open science”, and “open source”. Open washing is dominated by partial disclosure centered on weights, while open source combines shared weights with broad disclosure of training data sources, training code, and documentation that enables reproducibility. Easy access emphasizes hosted interfaces or packages, while internal artifacts remain limited. Open science prioritizes research reporting and archived materials over deployment convenience. Open weight occupies an intermediate position, with strong weight availability but uneven disclosure elsewhere. The segmentation is structured by three latent dimensions: “reproducibility” on axis one, “readiness” on axis two and “productization” on axis three. These results refine the common open weight versus open source dichotomy and support future work on the economic rationales behind each profile in contemporary model development and release.