When an AI company says its model is “open,” it usually means you can download it and run it on your own computer. That’s real, and it’s useful. It’s also not quite what “open source” has meant in software for the last few decades, and the two terms get used like they’re the same thing.
They overlap a lot. They aren’t the same.
Open Source
In regular software, open source means you get the code and you’re free to study it, change it, and share it without asking anyone. The Open Source Initiative (the group that’s kept that definition since the late ‘90s) published a version for AI in October 2024. To count, a model has to come with its weights, the code used to train it, and enough information about the training data that someone could rebuild something pretty similar.
Very few big models clear that bar. One that does is OLMo 3 from the nonprofit Ai2, released with its training data (about 9.3 trillion tokens of it), its code, and the snapshots it saved along the way. Researchers can check its homework.
Open Weights
Open weights means you get the finished model, the numbers it learned during training, and that’s about it. You can run it, fine-tune it, and build on it. You can’t see what it learned from or redo how it was made. That’s still a pretty big deal. Not long ago, you couldn’t download anything close to a top model.
Most “open” models are this. OpenAI’s gpt-oss models, released in August 2025, are open weights. So is Meta’s Llama, which the Open Source Initiative says fails the test for another reason too. Its license limits who can use it and how (really big companies need Meta’s permission first).
The tell
Ask whether you could rebuild it. If you get the weights, the training code, and a real description of the data, it’s open source. If you get the weights and a license to read, it’s open weights.
And if that license has a list of who isn’t allowed to use it, that’s open weights too, whatever the announcement calls it.
For most people running a model on a laptop, the difference doesn’t change much day to day, honestly. It matters more if you’re a researcher trying to figure out why a model does what it does, or a company trying to make sure the license won’t bite you later. That’s when “open” stops being a vibe and starts being a legal question.