Open-Weight vs. Closed Models: What "Open Source AI" Actually Means
Being able to download a model's weights is not the same thing as the model being open source in any traditional sense.
The phrase "open source AI" gets used loosely enough that it's worth pulling apart what it actually refers to, because the label is doing a lot less work than the traditional open source software definition would suggest.
When people call a model "open," they usually mean its weights — the enormous set of trained numeric parameters that determine how it behaves — are available for anyone to download and run. This is genuinely useful: you can run the model on your own hardware, fine-tune it, inspect its behavior, and use it without sending your data to someone else's servers or depending on their API staying available and affordably priced. Models like Meta's Llama family and Mistral's releases fall into this open-weight category.
What "open weights" usually does not include is the training data or the training code. In traditional open source software, you get the actual source — you can see exactly how it was built and, in principle, rebuild it yourself. With most open-weight AI models, you get the finished result of training without the recipe: not the dataset it learned from, not the exact training process, and often not enough detail to reproduce the model from scratch even if you had comparable computing resources. That gap is why some in the open source community argue these models shouldn't be called "open source" at all, and why organizations that maintain formal open source definitions have had to specifically address AI models as a distinct case.
Closed models sit at the other end: you interact with them only through an API or an app, the provider controls the weights entirely, and you have no ability to run the model yourself, inspect its internals, or use it without their infrastructure and pricing. In exchange, closed-model providers typically offer their most capable models, handle all the infrastructure and scaling, and update the model over time without you having to do anything.
The practical tradeoff for a developer is fairly direct. Open-weight models give you control, privacy, and the ability to run entirely offline or on your own infrastructure, at the cost of needing your own hardware, doing your own maintenance, and often trailing behind the most capable closed models on raw quality. Closed models generally offer stronger out-of-the-box capability and zero infrastructure burden, at the cost of ongoing API fees, dependence on a third party's uptime and pricing decisions, and sending your data to servers you don't control. Neither is universally correct — the choice depends on whether privacy and control or raw capability and convenience matter more for what you're building.