Open and closed models
In brief
A closed model is only available through its provider's service; an open model publishes its weights, which anyone can download and run. In between, licences and degrees of openness vary widely.
Definition
- A closed (or proprietary) model can only be used through its provider's app or API. Its parameters, called weights, stay secret.
- An open model (open weights) publishes its weights: you can download it, run it on your own machines and adapt it.
- A strictly open source model goes further: under the Open Source Initiative definition (2024), it must also give access to sufficient information about its training data and to the code needed to reproduce it. Few models meet all these conditions.
Why choose an open model
- Control over data: it can run on your servers or your computer, without sending information to a third party.
- Customisation: it can be adjusted (fine-tuning) to a specific need.
- Independence: no reliance on a single provider, or on its price or policy changes.
Why choose a closed model
- Often the most capable at the time of release.
- Nothing to install: an account is enough.
- Security, updates and infrastructure are handled by the provider.
Points to watch
- Read the licence: some "open" licences restrict commercial use or add conditions.
- Hardware cost: running a large open model requires powerful machines.
- Responsibility: in Europe, the AI Act sets different obligations depending on the type of model and its use.
Spotted a mistake or have an idea about this entry? Let us know