Some AI developers release model weights publicly so anyone can run, inspect, and modify them. Others keep weights private and provide access through controlled interfaces. The choice affects competition, security, safety, and who holds power over the technology.
Openness is a spectrum rather than a binary. Some releases include weights but not training data or code; some restrict commercial use through licensing; some publish everything. The label open is applied inconsistently across very different arrangements.
Released weights cannot be recalled. Safety measures built into an open model can generally be removed by anyone with modest resources and access to the file, which is the central asymmetry in the debate.
Closed models allow developers to monitor use, apply usage policies, and withdraw access. They also concentrate control over a consequential technology in a small number of firms, and limit independent verification of safety claims.