A licence attaches to a variant, not to a family name

Published 2026-09-22

The project registry holds one row per model a device may be offered, and every row carries a licence verdict rather than a licence name it hopes is close enough. The rule behind that is not caution for its own sake. It is the result of a defect that a green test suite did not catch.

One base, two licences

OpenGPT-X, the German project behind Teuken, publishes two instruct lines on the same 7B base. Their own model cards state the licences, and those cards are the primary source for the verdicts this project uses.

Teuken-7B-instruct-commercial-v0.4: the model card states Apache-2.0. In the registry this variant carries the verdict that may be routed.
Teuken-7B-instruct-v0.6: the model card states CC-BY-NC-4.0, a non-commercial licence. In the registry this variant is ineligible and the row exists to prove the gate refuses it.
Both rows state the same parameter size, because they are the same base. The size is the one thing the two have in common, and it is the one thing that says nothing about the licence.

The defect, and the direction it pointed

The registry originally recorded Teuken at family level as Apache-2.0 and commercially routable. A device announcing the bare family name could be holding either set of weights, and nothing in the row could tell them apart. The licence gate would have scheduled work under a commercial verdict that, for one of the two, was simply wrong.

The direction of that error is the part worth remembering. A verdict that is too strict loses a customer and someone complains. An over-permissive verdict schedules work that may not be lawful, nothing downstream refuses it, and nobody notices until a lawyer does. It surfaced only because a second registry, in a different repository, had recorded the other variant, which is a strange and lucky way for a licence defect to be found.

How the registry is written now

Each variant is pinned to its own model card, and the id names the variant rather than the family. A family-level identifier is not routable at all. The two rows that exist to be refused carry an empty list of execution tiers on purpose: no lane verified a tier for a model that must never be scheduled, and inventing one to make the table look complete would be a claim with nothing behind it.

The same trap in other rows

The pattern repeats across the reference list of EU-relevant models, and the shared guide these rows cite is where the verdicts come from.

Mistral Large 2 and Pixtral Large sit under the Mistral Research Licence, which the reference states is not commercially usable. Open weights and a permissive licence are not the same thing, even within the family of a single vendor.
Aleph Alpha Pharia-1 is research only, with commercial use handled by a contract rather than by the public download.
Meta Llama 4 uses the Meta Community Licence, which the reference records as excluding EU-domiciled companies from the multimodal models; for a European registry that is a refusal rather than a grant.
Most other rows carry no parameter count, context window, model size or memory floor at all. The reference does not state them, so the registry writes them as unverified. A plausible-looking number would have been the easy thing to type and the exact thing this project forbids.

What this changes for a reader building on it

Check the variant, not the brand, and check the model card rather than the family page. Licences change, and a version bump can change one. That is a one-line rule with an expensive counterexample, which is the best kind of documentation to have before the lawyer arrives.

Sources

Each link goes to the source the post cites. Where a source does not state a fact, the post shows it as unverified instead of filling it in.