AI inherited the moral glow of open-source software without inheriting a single agreed meaning of openness. In practice, “open” can describe several different layers—and access to one does not guarantee access to the others.
OPEN WEIGHT
Open-weight usually means the trained parameter files are available to download. You can often run and fine-tune the model, subject to its license. But the training data, data-cleaning methods, full training code, and reproducible recipe may remain unavailable.
DOWNLOADABLE IS A TECHNICAL CONDITION. OPEN IS A POLITICAL RELATION.
OPEN SOURCE
In ordinary software, open source means code is available under a license granting rights to use, study, modify, and redistribute it. For a model, source is harder to define: weights alone describe the artifact, not the process that produced it.
ASK FIVE QUESTIONS
Can I download the weights? Can I inspect the training code? Is the data named or available? May I modify and redistribute the result? Does the license restrict who can use it or for what? The answers matter more than the adjective on the launch post.