Open-source AI refers to the class of machine learning models whose weights, training data (partially or fully), and inference code are released under licenses permitting use, modification, and redistribution. Since 2023, the category has expanded from marginal curiosity to genuine alternative: Meta's Llama series, Mistral's releases, DeepSeek's 2025 models, and smaller-scale open weights from a growing ecosystem of research labs and independent developers have demonstrated that frontier or near-frontier capability is achievable outside the proprietary platforms. In Jacobs's framework, this ecosystem is the mechanism by which the supply of cheap creative space can be renewed against the inevitable pressure of rent increases.
The structural significance of open-source AI is not primarily technical. It is economic. A digital economy entirely dependent on a handful of proprietary model providers is a digital economy whose landlords can raise rents at any time, change terms at any time, and demolish the buildings where experimentation happens. Open-source