Three days ago we published a full day of MiniMax H3 tests and told you we could not put any of it in a video, because of the licence. On Friday we applied for authorization. This morning it came back approved.

Before the useful part, the correction.

What we got wrong

On 30 August we wrote that the H3 licence “puts an application form in front of anyone in the US, EU, UK or South Korea” and called it “a gate rather than a ban”. We even framed it as a commercial-use problem, and we described that as a correction of an earlier, harsher reading of ours.

The harsher reading was closer. When we went through it properly to apply, the relevant clause does not restrict commercial use in those four territories. It withholds the grant entirely. In an excluded region you do not have a limited licence, you have no licence, whether you are selling something or posting it for free. There is a separate path where excluded parties can apply to the maker for authorization, which is the part we got right, and it is the reason any of this ended well.

So we over-corrected in public and then repeated the softer version in a guide. That is on us. Both pieces now carry a dated note pointing here.

What applying actually took

A short form on the maker’s site. A written commitment to the content safeguards we already run, which for us meant describing the gates that keep our automated pipeline from publishing things it should not. An agreement to credit the model on screen wherever we use it. And a clause we now treat as law.

From submission to approval was about three days. The authorization is issued to our business, and it is conditioned on us continuing to do what we said we would do in the request.

We are not going to pretend this makes us licence experts. We read a document, filled in a form, and got an answer. That is the whole of our expertise here.

The clause that should stop this hobby cold

Nothing the model generates may ever be used to train or improve another model.

Read that again if you are in this community, because harvesting your own generations into a training set is one of the most normal things people do. You render a few hundred images of a character you like and you train a LoRA on them. It is a default habit. On this model it is a licence breach.

It is our habit too, which is exactly why we flagged it internally the moment the weights landed rather than after. Our pipeline routinely turns generated frames into training data for the next thing, so H3 output now lives in folders tagged so that nobody, human or agent, walks it into a dataset by accident.

If you take one thing from this piece, take that one. It is not enforced by anything technical. Nothing will stop you. It is just a term you agreed to when you downloaded the weights.

What changes, and what does not

For us: the bench we published on Friday is no longer bench-only. H3 can appear in what we make, with the model credited on screen, and with that no-training clause treated as an absolute. The full write up of what it can do is unchanged on the technical side, because measurements do not care about paperwork.

For you: nothing. Our authorization covers us, not our readers. If you are in one of the excluded territories and you want to use this model for anything, you need your own, and the form is on the maker’s platform site.

The broader pattern is worth naming, because it is going to keep happening. Open weights and an open licence are not the same thing, and a growing number of releases are open conditionally: the file downloads anywhere, and the licence carries territory limits, branding requirements, revenue caps or field of use restrictions that only bite once your project matters. Our two minute licence routine is the check, and the calculator now says plainly that we only recommend models we can legally run ourselves.

There is one small satisfaction in this. We spent two articles telling people to read the licence before building on a model. Then we read one properly, found we had described it wrong, filled in the form, and got a yes in three days. The advice held up. We were just sloppier at following it than we were at giving it.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.