A model whose trained numbers (weights) are published so anyone can download them.
The heart of an AI model is a huge set of numbers fixed during training, called weights. An open-weight model publishes those weights as files, so anyone can download the model, run it on their own machine and adjust it. A closed model stays on the company's servers; you can only talk to it through an app.
How "open" a model is varies. Many release the weights but not the training data or the full recipe, and some attach rules about how you may use them. That's why people argue about whether these models should be called "open source".
Think of it like this
Imagine one restaurant that only delivers meals, and another that hands you a whole jar of its secret sauce. With the jar, you can take it home and cook with it your own way. Open weights are the jar, but you often don't get the recipe that says what went into it.
For example
Download the weight files and, if your laptop is powerful enough, you can run the model with no internet at all.
Common mix-ups
Mix-upAn open-weights model is open source, recipe and all.
ActuallyIn 2024 the OSI, the group that looks after the open-source definition, said an AI counts as open source only if three things are released: information on the training data, the code used to build it, and the weights. By that standard, most well-known released models, such as Llama, Gemma and Qwen, don't qualify.
Mix-upIf a model is released, you can use it however you like.
ActuallyReleased models can come with conditions. LLaMA 1 was for non-commercial research by application (record E0195); Llama 2 allowed commercial use under its own licence (record E0247).
Try it in class
For teachers at the start of a lesson — or for students working on their own.
- Question
- What must be released before a model deserves to be called "open"?
- Activity
- Each group makes a table with columns for weights, training-data information, code and conditions of use. Read records E0195, E0247 and E0455 and fill in released, partly or not said for each, then agree on one sentence that states your group's test for an "open model".
Records about this
Source-checked moments from AI history.
Established the grounds for ruling that most of Llama, Gemma and Qwen are not open source
Popularized open weights, and the starting point of the debate over whether it is open source
The first large attempt to distribute frontier-class weights after vetting