How do open-source and open-weight models differ?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
An open-weight model makes its trained numerical settings available for others to obtain and use under specified terms. Those settings, often called weights or parameters, are learned during training and help determine the model’s responses. Downloading them can allow a model to run outside the developer’s own service, provided suitable software and hardware are available.
Open-source AI is a broader claim. Under the Open Source Initiative’s Open Source AI Definition 1.0, the system must support freedoms to use, study, modify and share, and provide the required code, parameters and information about training data. The definition does not simply equate downloadable weights with open-source status.
Having the trained weights lets you use—and, where the license allows, adapt—the model. Studying or reproducing the training process also requires the relevant code and information about the data. These are different levels of access. The Open Source AI Definition specifies what must be supplied; it is not simply a request for more transparency.
Training data information also needs careful wording. The definition requires detailed information about the data and access arrangements; it does not mean that every item of private or otherwise unshareable training data must be publicly disclosed. The linked definition explains the conditions. Different organizations sometimes use “open” more loosely, so naming the standard avoids ambiguity.
A license is the set of terms governing permitted use and redistribution. Check the actual terms for the particular release. “Available to download” is a statement about access, not a complete account of what can legally or practically be done with the model.
Open, free and local are separate properties. An open-weight model may be offered through a paid hosting service. Running it locally means performing the model’s calculations on your own hardware. That may require substantial memory and technical work, and connected tools can still send information elsewhere.
A university research team might choose downloadable weights so it can hold a model version fixed while comparing methods for extracting information from clinical trial reports. That provides version control, but the team still needs to understand the license, supporting software and reproducibility limits. A hospital considering local use faces additional questions about maintenance, access and every route taken by the data. Open weights, open source and local execution answer different questions; none establishes that the resulting healthcare application performs well.
Sources: Open Source Initiative — Open Source AI Definition 1.0 and FAQ.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy