Using a model commercially
You want to ship a product built on someone else's open model, and you want to be straight about it. This page is the whole job: work out whether you owe anything, compute what each creator is owed, and pay them. It happens once a quarter, and nothing is filed with OMLA — there is no account to open, no report to submit, and no one to ask for permission.
First: do you owe anything?
Most people reading this do not. Work down the list and stop at the first one that fits.
Free — non-commercial use
Published academic research at any size, personal projects, and internal training, evaluation and development that produces no revenue. Nothing to compute, nothing to pay, nothing to tell anyone.
Free — commercial, but under the revenue floor
Your entity is under $200,000 in annual revenue and your product does not directly sell the model's outputs. Startups and small shops fall here. Still free.
You owe — above the floor, or selling outputs directly
Either your entity is over the $200K floor, or you are directly monetizing the model's outputs at any size (selling generated images, charging per generation, reselling inference). Now you pick one of the two options below.
Two ways to settle
| Option | What it means | What it costs |
|---|---|---|
| Copyleft | Release the weights of your derivative under the same license, so the next person gets what you got. | No money. You give up exclusivity on the derivative. |
| Pay | Keep your derivative closed and pay the creators instead. | 30% of the greater of attributable revenue or attributable model run cost, split across every model involved. |
The run-cost floor exists so "our AI feature is free, we made $0 from it" is not an answer — if you would have paid an API provider for equivalent inference, that is the floor. Full text: License §5, Copyleft or Pay.
What you actually do, once a quarter
Exactly one calendar quarter at a time (2026-Q1 through 2026-Q4), settled within 60 days of quarter end.
- List the models. Which OMLA-licensed models did your product run this quarter, and how many generations did each combination serve? Your inference logs already know this.
- Get each model's signed manifest — and its parents' too. They ship next to the weights, or come from the registry. Verify the signature — that is what tells you the payment details were not tampered with. If a model is a fine-tune, you need the manifest of what it was tuned from, and so on up the chain; missing one sends money to the wrong creator.
- Work out the pot. 30% of the greater of attributable revenue or attributable run cost, in integer cents.
- Run the resolver. Paste it all into the calculator below, or run the resolver yourself in your own build. It is the same code.
- Pay each wallet directly at the pointers in the manifest, and keep the settlement sheet in your own records. Send nothing to OMLA.
The settlement calculator
Runs entirely in this tab against the same open resolver the license points at — no network calls, nothing uploaded, no OMLA server involved. Refreshing the page throws it all away.
1 Add every model you used
Paste one signed manifest at a time. Include the parents too — if you use a fine-tune, the royalty flows up its lineage, and the calculator can only follow that chain if you give it the parent manifests as well.
2 Describe how they were used
A combination is a set of models that ran together to produce one output — a base checkpoint plus whichever adapters were active. Add one row per distinct combination, with how many generations it served this quarter. If you only ever ran one model, that is one row.
3 Enter the money
What equivalent API inference would have cost. The royalty is charged on whichever of the two is larger.
Verify a manifest
Before you trust a manifest's payment details, check that the creator actually signed them. This verifies the Ed25519 signature over the whole document — identity, lineage, split, certifications and payment pointers. If someone edited a payment address after the fact, this fails.
What if the model you use has no manifest?
This is the common case today, and pretending otherwise would waste your time. Most open models on the internet have never heard of OMLA, and the registry is young. Here is the honest position:
- No OMLA manifest means no OMLA obligation. The license is what creates the duty to pay. A model published under Apache-2.0, MIT, a Llama Community License or anything else is governed by that license, not this one. OMLA does not claim a royalty on models that never opted in, and never will.
- You cannot owe money to someone who has not said where to send it. There is no escrow, no holding account, no "pay OMLA and we'll find them later." That is the no-custody rule, and it is permanent.
- If you want to pay a creator who has not published a manifest, point them at the publish page. It takes a few minutes and produces a file they can drop next to their weights. You get a clean licensing position; they get paid.
- If a manifest exists but you are missing its parents, stop and go find them. The resolver is deliberately total: rather than leaving a gap, it folds an unresolvable share pro-rata into the creators it can resolve. The arithmetic still balances, so nothing looks wrong — but the money goes to the wrong person, and a fine-tune that should keep 5% quietly collects the foundation creator's 95%. The calculator checks for this before computing and refuses to let it pass silently.
The practical shape of this: OMLA is worth using when the models you depend on carry manifests. Until one does, this page costs you nothing and asks you for nothing.
What OMLA will never ask you for
- An account, a login, or an API key.
- Your revenue, your usage numbers, or your customer data.
- A royalty report, a settlement sheet, or proof of payment.
- Money. OMLA takes no cut and never touches the funds — payment is wallet to wallet, you to the creator.
If anything ever asks you for those in OMLA's name, it is not OMLA. The complete list of what the organization does is on the about page.