Open weights vs open source
- Training
- Standard term
Open weights means you can download a model's trained numbers, its weights, under its license. Open-source AI also shares what you need to study and rebuild the model, with terms allowing use, change, and sharing.
Running a model, or training it further on your own examples (fine-tuning), needs only its weights. The weights are the end product of training, though, and hold no record of the code or the data that produced them, which developers often keep private.
Where you’ll meet it
You want a model that runs on your own laptop, so client documents never leave it. You find one whose page calls it open. The download includes the model files and a license, no training code, and a one-line note about the training data.
What to do
Read it as an open-weights release, and for your plan that is enough. Before you use it on paid client work, open the license and find what it says about commercial use.
Easy to mix up
With an API, the provider runs the model on its own servers and you send it requests. With open weights, you download the model and run it on a computer you choose.
Check yourself
A lab publishes its training code and a full description of its data under open terms, but keeps the trained weights private. Does the model count as open-source AI?
Go deeperThe Open Source Initiative drew the line in 2024
The Open Source Initiative, which defines open-source software, published its Open Source AI Definition in October 2024. It keeps the freedoms of its software definition, and its test for the data description is whether a skilled person could use it to build a substantially equivalent system. Critics object that it doesn't require releasing the data itself.
Licenses on popular weights releases often add conditions that open-source licenses rule out, such as a cap on how many users a product may have or a list of banned uses. Conditions like these are why many models described as open fall short of the definition.