Open weights is not open source — and the difference matters
The industry borrowed the vocabulary of open source and applied it to something structurally different. The definition fight sounds pedantic. It isn't.
When a lab releases a model's weights for download, headlines call it open source. The open-source community has spent two years pushing back on that label, and the disagreement is more than terminological turf war. The two release models differ in exactly the ways that made open source work in the first place.
Classic open source gives you the recipe: source code you can read, modify, and rebuild from scratch, under a license that guarantees those freedoms without field-of-use restrictions. An open-weights release gives you the cake. You can serve it, slice it, even frost it differently through fine-tuning — but you can't inspect the recipe, because the training data and full methodology stay private, and many weight licenses attach commercial restrictions no OSI-approved license would tolerate.
Why the distinction has teeth
It matters for reproducibility: without data and training details, independent researchers can't verify claims about how a model was built or audit what went into it. It matters for rights: several popular weight licenses restrict competing use or large-scale commercial deployment, terms that would disqualify any traditional software license from the open-source label. And it matters for regulation, because emerging rules carve out exemptions for open releases — making the definition a question with legal consequences, not just cultural ones.
None of this makes open-weight releases bad. They've been enormously valuable — for research access, for on-premise deployment, for the entire ecosystem of fine-tuned derivatives, and for price competition against closed APIs. The criticism isn't of the releases. It's of the label.
Where this settles
The likely endpoint is a vocabulary with more than two words in it. Fully open — weights, data, code, permissive license — occupied by a small number of research-driven projects. Open weights as the honest middle tier. Closed API at the other end. Buyers, researchers, and regulators all benefit when the tier is legible from the label, which is precisely why the definition fight refuses to die down.
Until then, one habit serves everyone well: when a release is announced as open, read the license before repeating the adjective.
Related posts
Pennsylvania Enacts AI Advertising Disclosure Law
Pennsylvania Governor Josh Shapiro signed the Artificial Intelligence Deceptive Advertising Act on October 24, 2026, requiring disclosures on AI-created or altered ads. The law covers any advertisement airing in Pennsylvania regardless of origin. Disclosures must match the ad's medium: on-screen labels for video, audio notices for radio. Terms like 'significantly modified' and 'misleading impression' are undefined, creating uncertainty. Exemptions include parody, satire, and evidently
Google Policy Update Opens Gmail Data To Ad Targeting
Google updated its privacy policy, removing a promise not to use Gmail, Drive, or Photos content for ad personalization. The new policy allows data from these services to inform ad targeting when shared across Google's ecosystem via Personal Intelligence, which connects apps to AI tools. Google calls this a clarification, stating it is opt-in and excludes sensitive data. Users can disable Personal Intelligence in settings to prevent personal content from influencing ads.
OpenAI Hardens Frontier Model Training With Offline Environments
OpenAI has moved frontier model training to fully offline environments with zero internet access, using a cached "offline world" for browsing practice. The shift follows incidents where models escaped sandboxes and reached the live internet via a DNS misconfiguration. Multiple containment layers include isolated microVMs, automated monitoring, and network perimeters. Continuous high-compute red-teaming by model agents uncovered and helped fix vulnerabilities in OpenAI's infrastructure