Read the conditions

Open weights describes a distribution method, not a permission set. The industry keeps using one word for two very different licences.

Apache 2.0 and MIT permit nearly unrestricted commercial use. Llama, Grok and MiniMax ship community licences with conditions. Both get called open.

The gap is not cosmetic

User thresholds. Field-of-use restrictions. Naming requirements. Clauses that terminate on litigation. All of these appear in community licences and none appear in Apache 2.0. Whether they bind you depends entirely on what you are building.

Most users are unaffected is a different claim from open, and the industry keeps using the second word for the first situation.

How it actually fails

Late discovery. A team evaluates on capability, builds the prototype, then finds at legal review that the terms do not permit the deployment they designed for. Weeks gone, because the licence was treated as paperwork rather than as a specification.

The month's biggest open release carries exactly this shape — a full-modal video model, weights public, community terms.

None of this makes the labs wrong

A company spending heavily on a model has legitimate reasons to stop competitors reselling it. The objection is not to the conditions, it is to the vocabulary that hides them.

The infrastructure went the other way

Worth contrasting with a protocol that reached a billion downloads without a consortium or a fight. Standards emerged there because no vendor could defend the integration layer alone. Nobody has that problem with weights.

What this asks of you

Read the licence at evaluation time, not at review time, and write down the specific clause that permits your deployment. Twenty minutes, and it is the difference between shipping and re-architecting.

And keep two questions apart in your own policy. Can we use it is answered. Can we inspect what it was trained on is not, and no licence in the field answers it.

OpenCurious — Top 33 Open-Source AI Models (2026): The Complete Table → · Pinggy — Best Video Generation AI Models in 2026 →