Open-weight video licenses and what you can ship
Open weights do not mean open usage. Four clauses decide whether a downloaded video model can run a paid campaign — and which entity has to pass the test.
"Open weights" is a distribution fact, not a permission. It tells you the trained model file is downloadable. It tells you nothing about whether you can run a client's Black Friday campaign on it, whether your agency's revenue puts you on the wrong side of a threshold, or whether the model's name has to appear somewhere in the credits. Those answers live in a license document that most people skim after they've already downloaded the checkpoint, which is the wrong order.
The category is genuinely wide. Alibaba's Wan 2.7 shipped in April 2026 under Apache 2.0 — about as permissive as a real license gets. Lightricks put LTX-2.5 out on August 11, 2026 under the LTX-2.x Community License, which is free for commercial use under a $10M annual revenue ceiling with a paid license required above it. Ideogram published inference code and weights for Ideogram 4 on June 3, 2026, plus quantized fp8 and nf4 builds on Hugging Face. Same two words on the announcement, three different deals. The open-weights wave post covers the releases; this one covers the paperwork.
Clause 1: the commercial-use condition
The first thing to find, and the one that most often turns out to be conditional rather than absent.
Three shapes recur. Unconditional — Apache 2.0 and similar permit commercial use with no revenue test, no field-of-use restriction and no separate agreement. Threshold-gated — commercial use is free below a stated size and requires a paid license above it, which is the LTX-2.x Community License shape. Research-only — no commercial use at all, which is rarer in video but does exist and is the one that quietly kills a project three weeks in.
The trap in the threshold shape isn't the number, it's the entity the number applies to. A $10M annual revenue ceiling is a test on a company, not on a campaign or a project. If you're a two-person studio, you pass and it's irrelevant. If you're a contractor running inference on your own hardware for a client above that line, you need to know whether the license tests the entity running the model or the entity benefiting from the output — those are different questions and licenses answer them differently. Read the definition of "you" in the license, not just the threshold.
Clause 2: territory exclusion
The clause most people don't know to look for, because it doesn't exist in the licenses they're used to. Some community licenses in this category carve out jurisdictions — the license grant simply doesn't extend to users or use in named territories.
This matters in two directions and the second one is the surprise. Obviously it matters if you're based somewhere excluded. Less obviously, it matters if your client, your client's audience, or your distribution is. A campaign generated in a permitted territory and served into an excluded one is a question the license may or may not answer cleanly, and "may or may not" is not a position to be in after the media is booked.
Practical rule: if the license names any jurisdiction at all, stop and read that section in full before you download anything. Territory language is never boilerplate — if it's there, somebody put it there on purpose.
Clause 3: who owns the outputs
The clause that trips up the most paid work, because the intuition is wrong. A weights license governs the weights. Whether it reaches the generations is a separate question, and licenses handle it in at least three ways: some are explicitly silent, some disclaim any ownership claim over outputs, and some attach conditions to outputs that mirror the conditions on the model.
Three things to establish before a paid campaign:
- Does the license assert any interest in generated output at all? If it's silent, you're relying on general law rather than a grant, which is a defensible position but not the same as a written one.
- Do output conditions travel to your client? If the license restricts what outputs may be used for, and you hand the files to a client who then does something else with them, the restriction didn't stop existing.
- Is there an acceptable-use policy attached separately from the license? Many releases split the two. The AUP is where the categorical bans live, and it usually binds output use, not just model use.
Usage rights is the concept to hold onto here: the right to generate and the right to publish are separate grants, and an open-weight release frequently gives you the first without saying much about the second.
Clause 4: attribution and naming
The cheapest clause to comply with and the easiest to breach by omission.
Apache 2.0 — the Wan 2.7 case — is permissive but not obligation-free. It requires you to retain copyright notices and the license text in redistributions, include the NOTICE file if one is provided, and state that you changed the files if you modified them. For a studio running inference internally and shipping finished MP4s, most of that never triggers, because you aren't redistributing the model. The moment you ship a product, a plugin, a container image or a fine-tuned checkpoint that contains the weights, it does.
The other half is naming. Some community licenses require derivative models to carry the base model's name, or require the model to be credited in materials describing the product. That's a brand decision disguised as a legal one, and it's worth catching before a client asks why a third party's name is in the deliverable.
| Clause | Where it usually lives | Does it block a paid campaign? |
|---|---|---|
| Commercial-use condition | Section 2 of the license | Yes, if research-only or if you're over the threshold and unlicensed |
| Territory exclusion | Definitions or a standalone section | Yes, for excluded users or use |
| Output ownership / AUP | The license, plus a separate AUP | Sometimes — usually restricts categories rather than blocking outright |
| Attribution and naming | Redistribution section | No, but it creates a deliverable obligation |
The distinction that resolves most of this, and the pre-flight
Whichever license applies to the downloaded weights, it does not follow the model onto somebody else's hosted endpoint. If you call a model through a hosted API, the terms governing that call are the host's terms. The open-weight license is the deal you get for running it yourself, on your own hardware, on your own account.
That's the actual fork in the decision. Self-hosting hands you the weights and hands you the license compliance work — the entity test, the territory read, the AUP, the redistribution obligations if you ever ship the model rather than its outputs. A hosted route hands you one commercial agreement with the host and no checkpoint to manage. Neither is strictly better, and running open-weight video on your own terms works through when self-hosting actually pays.
On the hosted side, Versely runs open-lineage models alongside closed ones with no distinction at the point of use. Wan 2.7 Text to Video bills per second of output with separate SD/HD and 4K rates — 10 credits per second at SD/HD and 15 at 4K — and the same is true across the Wan family. The Apache 2.0 license on Alibaba's release is what made those weights broadly available in the first place; it isn't the agreement that governs your generation when the inference happens on a hosted endpoint. Both facts are true at once and people routinely collapse them into one.
So run this pre-flight before the first generation, not before the delivery deadline.
- Name the license file. Not "it's open" — the actual file, the actual version.
- Find the definition of "you." Then decide which entity in your chain has to pass any test.
- Search the document for jurisdiction names. If any appear, read that section end to end.
- Locate the output clause, or confirm there isn't one. Write down which it is.
- Find the acceptable-use policy. It's often a separate URL and it's often stricter than the license.
- Decide whether you are redistributing anything. Shipping outputs is not redistribution; shipping the model, a fine-tune or a container that contains it is.
- Record all of it in the project file. Licensing for business AI content covers keeping that record in a form that survives a client audit.
FAQ
Does an open-weight license mean the model is open source?
No. Open weights means the trained parameters are published and runnable; open source in the full sense would also mean the training code and data. You get the artifact, not the recipe. Some open-weight releases do use genuine open-source licenses on the weights — Apache 2.0 on Wan 2.7 is one — but the two claims are independent and one doesn't imply the other. The LTX-2.5 release post works through that distinction on a specific model.
If I fine-tune an open-weight model, whose license applies to the result?
The base model's, almost always, plus whatever obligations attach to derivatives specifically. Fine-tuning does not launder a license — a derivative of a threshold-gated model is generally still threshold-gated, and a derivative of a model with a naming requirement generally still carries the naming requirement. This is the case where the redistribution clauses stop being theoretical, because a fine-tuned checkpoint is a thing you might actually hand to someone.
Do I need to disclose which model generated a client's campaign?
Licensing and disclosure are separate obligations that people merge. An attribution clause is a contractual requirement from the model's publisher; a disclosure duty is about telling an audience the content is synthetic. Satisfying one does nothing for the other. Decide them independently and write both positions down.
Is a hosted model with an open-weight lineage governed by the open license?
No. The license attached to the weights governs your use of the weights. When you generate through a hosted endpoint, the host's commercial terms govern that generation. The lineage is useful context and it's not the agreement you're operating under.