There are three primary categories of artificial intelligence models are commonly distinguished: “open weight” models, “open source” models, and “closed” models. It is necessary to delineate their fundamental characteristics.
An “open weight” model is defined as a downloadable and executable model, deployable on personal or cloud infrastructure, yet without disclosure of its design mechanisms. This category includes, for instance, models such as Llama, Qwen, or certain versions of Mistral.
The coexistence of these three paradigms addresses distinct needs for both users and providers.
“Open source” models offer a higher degree of disclosure. Not only are they downloadable, but they also provide comprehensive access to their components, thereby enabling analysis, comprehension, and retraining. This access encompasses the complete training data as well as detailed technical specifications. Illustrative examples include the BLOOM models or, more recently, OLMo.
In opposition, “closed” models are not downloadable. Their use is restricted to dedicated online services. Models such as GPT-4, GPT-5, or Gemini are typical examples. Their internal architecture and development processes remain opaque. These systems frequently incorporate additional layers designed to enhance their robustness and functionality, such as security filters or internet access modules.
The coexistence of these three paradigms addresses distinct needs for both users and providers.
What control do users have?
From the user’s perspective, the central issue is one of control. Some prioritize the autonomous deployment of models within secure infrastructures, while others require the ability to customize or modify them.
For providers, closed and open weight models serve to preserve intellectual property and technical know-how. Closed models further provide increased control over usage conditions, notably through query filtering. The open weight strategy, on the other hand, fosters broad adoption and the development of an economic ecosystem around the model.
Finally, open-source models address a cross-cutting imperative of transparency and trust. Although their performance may be comparatively limited, they constitute essential resources for academic research, regulatory bodies, and any organization tasked with explaining the underlying mechanisms of a model’s operation.
Each typology possesses its specific legitimacy and utility. None is superior in an absolute sense; the selection must stem from a rigorous analysis of requirements, supplemented by a careful examination of the licenses and terms of use established by providers.
