In brief
With an open-source AI model, the focus is on the freedom to use, examine, adapt and share the system. Simply offering a free chatbot or downloadable model files does not, in itself, make a model fully open source. In discussions about AI, the term is used broadly; therefore, look at what is actually available and what the licence permits.
What are open weights?
‘Open weights’ means that the trained model parameters are available. This may allow you to run and customise the model yourself. However, the licence may contain restrictions, for example regarding certain applications or distribution. Furthermore, ‘open weights’ does not automatically imply that the training code and sufficient information about the training data are available. The OSI definition sets out broader requirements for an open AI system.
How does this differ from a closed API?
With a closed API, you send a request to a service that runs the model for you. You receive the response, but usually not the weights or the ability to manage the same model version independently. A provider may also offer an API for a model with weights available. Access to files and access to a service are therefore different matters.
A practical example
An organisation wishes to make its internal manuals searchable. With a suitable model and appropriate licence, it can set up the processing on its own server. This also involves a search index, access rights and a user interface. The model download is one component of that solution. Managing these components takes time, even if the download itself is free of charge.
‘Open’ does not automatically mean secure or free
Available weights do not provide any definitive information about answer quality, the origin of training data or the security of your installation. Running the system yourself entails costs for hardware, electricity and management. You should also check whether the licence terms are suitable for your application. A local model with online search tools may still send data to external services.
What does this mean for your choice?
For a selected model, retain the version, licence and configuration. Test with your own tasks and clarify which components you can replace independently. In that case, openness becomes a practical way to maintain control, rather than just an attractive label. The choice between running the system yourself and using a service remains dependent on capacity, knowledge and usage.
From concept to practice
How AITJE applies this
For a custom AI environment, we assess which model files and usage rights are available. We select an implementation that is compatible with the hardware and test the required tasks. In doing so, we specify the model version and licence. We do not automatically classify downloadable weights as fully open source: the rights and available components determine what you can manage and customise independently.
Installation and setupFrom explanation to application
Would you like to know what this means for your work?
