"Open" is not a licence

Eleven of thirty models described as open carry real commercial conditions. The word has become a marketing category with no legal content, and the models people actually deploy are disproportionately the ones with terms attached.

An August audit read the licences on thirty models publicly described as open, open-weight or open-source. Seventeen were genuinely permissive. Eleven carried a real commercial condition — user thresholds, attribution requirements, field-of-use bans, regional exclusions.

Two in three is not a reliable signal

If the label were wrong occasionally, it would be a caveat. At more than a third, it carries no information. "Open-weight" tells you the weights are downloadable and nothing whatsoever about whether you may ship a product on them.

And the conditions found are not technicalities. A user threshold changes the licence at the exact moment the product works. A field-of-use ban forecloses an application no matter how well the model performs. A regional exclusion makes the same deployment lawful in one market and not in the next.

The distinction people keep collapsing

Open weights: the finished model, to run and fine-tune. Open source: the model plus the training code and data information needed to rebuild and audit it. Almost everything shipping under the open banner is the first. The second is rare enough that when it appears it is the story.

This matters beyond pedantry, because the auditability that makes open source trustworthy is exactly the property open weights do not provide. You can inspect behaviour; you cannot inspect provenance.

Where the conditions cluster

The leading open-weight models are predominantly from Chinese labs — Qwen3.8-Max at the top of the August ranking, Kimi K3 reported fourth globally across all models. Regional exclusion clauses read rather differently when licensor and deployer sit in different jurisdictions, and cross-border enforcement of a licence term is nobody’s idea of a solved problem.

The rule that would prevent all of this

Read the licence attached to the checkpoint you are downloading. Not the family's reputation, not the announcement post, not what the aggregator's comparison table says. Licences travel with artefacts; reputations travel with names; and they come apart most often on the models everyone is already 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) → · BenchLM — Open-Source LLM Leaderboard 2026: 101 Models Ranked →