What is the difference between open weights, open source, and proprietary models?

People use the word “open” loosely, but for AI models it covers three quite different things. The honest way to read a model release is to ask what was actually handed over.

A model’s weights are the numbers learned during training. They are what the model “knows,” and they are what you run when you generate text. An open-weights release means those trained parameters are published for you to download, run, and build on. Most well-known “open” models, including Meta’s Llama family, are open weights.^[3] What an open-weights release usually does not include is the training data or the code used to create the model.^[2]

Open source AI is a higher bar. The Open Source Initiative’s definition asks for the freedom to use, study, modify, and share a system, and it lists three things a release should make available: detailed information about the training data, the complete source code used to train and run the system, and the model parameters.^[1] The point of this extra material is reproducibility. Without the data and code, others cannot inspect how the model was built or recreate it.^[2]

A proprietary model is the closed end of the spectrum. You reach it through an API, you do not get the weights, and you cannot run it on your own hardware. You are buying access to the output, not the model itself.