An audit of thirty "open" models found eleven with commercial conditions
A census of models publicly described as open, open-weight or open-source found 17 of 30 genuinely permissive and 11 of 30 carrying a real commercial condition. The word on the announcement is not the word in the licence file.
An August audit read the licences on thirty models publicly described as "open", "open-weight" or "open-source". Seventeen were genuinely permissive — MIT, Apache 2.0 or equivalent. Eleven carried a real commercial condition: user thresholds, attribution requirements, field-of-use bans or regional exclusions.
More than a third is not an edge case
If one model in ten had a surprise in its licence, that would be a story about that model. Eleven in thirty means the category label carries no reliable information, and anyone treating "open-weight" as a synonym for "you may deploy this" is right about two-thirds of the time.
The conditions found are not decorative. A user threshold means the licence changes exactly when the product succeeds. A field-of-use ban means an entire application is prohibited regardless of how well the model performs. A regional exclusion means the deployment is lawful in one market and not another.
The distinction underneath
Open weights give you the finished model to run and fine-tune. Open source gives you the model plus the training code and the data information needed to rebuild and audit it. Almost everything shipping under the open banner is the former, and the two get used interchangeably by people who know better as well as people who do not.
What to do about it
Read the licence attached to the specific checkpoint, not the family, and not the blog post announcing it. The licence travels with the artefact; the reputation travels with the name; they come apart quietly and they come apart most often on the models everybody is using.
Digital Applied — We Read the Licences on 2026 Open-Weight Models → · Cooley — Unlocking the Weights: What Enterprises Should Know Before Deploying Open-Weight AI Models → · GEO Toolbox — Open Weights vs Open Source: The Real Difference (2026) →