AI News
Open-Weight vs Open-Source AI: What's the Real Difference?
Every few weeks, a headline announces a new "open-source" AI model — and every few weeks, someone points out that it isn't open source at all. The confusion is constant because the two terms sound alike and the industry uses them loosely. The distinction is simple, and it matters if you build with these models.
What "open weights" means
A model's weights are the vast set of numbers learned during training — the actual artifact that makes the model behave the way it does. An "open-weight" release publishes those numbers for download. With the weights, you can run the model on your own hardware, fine-tune it on your own data, inspect its behavior, and deploy it without depending on the original vendor's servers.
That is genuinely valuable. It means lower costs, more privacy, no per-token billing, and the ability to audit exactly what you are running. Most of the celebrated "open" releases of the last few years — downloadable from hubs like Hugging Face — are open-weight releases.
What "open source" properly means
"Open source" has a precise definition, maintained for decades by the Open Source Initiative. Software is open source when its license gives everyone the freedom to use, study, modify, and share it — for any purpose, without discrimination. The key phrase is "for any purpose": a license that says you may use the software except for certain commercial uses, or only below a certain user count, is not an open-source license, no matter what the announcement says.
Applied to AI, true open source would mean the weights, and ideally the training code and data information, released under terms with no field-of-use restrictions. Very few prominent model releases meet that bar.
Why the confusion persists
Two reasons. First, marketing: "open source" sounds better than "open weight," and vendors know it. Second, many widely used model licenses sit in the middle — the weights are downloadable and usable, including commercially, but with conditions attached, such as restrictions above certain scale or requirements to share modifications. These are real, useful freedoms, but they are not the freedoms the term "open source" promises.
The result is a recurring cycle: a launch announcement says "open," the community celebrates, and then someone reads the license and the argument starts. If you have seen that argument, you have seen this distinction in action.
Why it matters to you
If you are a developer choosing a model, the license determines what you are allowed to build. Can you use it in a commercial product? Can you fine-tune it and keep the result private? Can you deploy it for a client? The answers live in the license text, not the headline. A model with restrictive terms can still be the right choice — but you should know the terms before you build on it.
If you are following AI policy, the distinction matters because "open" is doing political work: it shapes debates about safety, competition, and who controls the technology. Precision here is not pedantry; it is the difference between a real check on concentrated power and a marketing claim.
How to read a model release
Ignore the headline and read three things: the license (what uses are permitted and restricted), what exactly is released (weights only, or also training code and data details), and who can enforce the terms against you. A release with downloadable weights and a permissive license is close to the open-source ideal. A release with downloadable weights and a long list of conditions is open-weight — useful, but different. Call each one what it is, and you will never be confused by the next launch announcement.