Choosing a licence for an open-weight model is a governance decision, not a download decision. The licence, not the fact that the weights are publicly available, determines whether you can use the model commercially, what you may build on top of it, and what you must disclose to your own users. A 2026 Cooley alert for enterprises frames licensing as the threshold issue when evaluating open-weight deployment, and a practitioner playbook from Rodan advises teams to check the licence before engineers start work and to record the result where the governance team can find it.
This guide is written for technical leads, procurement staff, and in-house counsel assessing a self-hosted model for a commercial product or an internal workflow. It gives you a repeatable decision procedure, the licence terms that most often change the answer, and the records worth keeping. It is not legal advice; both sources recommend that a qualified adviser confirm the final decision.
Open weight is not open source
The distinction matters because it changes what you can assume. Open-source software is generally distributed under standardized, well-understood licences that grant broad rights to use, modify, and redistribute the code, as Cooley notes. Open-weight models work differently: they make a model’s weights, the numerical settings a model learns during training that shape how it responds, available for download and local deployment. They still require integration through a separate software platform to run and manage the model, and they often remain subject to bespoke contractual terms covering commercial use, intellectual property, redistribution, attribution, and downstream deployment.
Rodan makes the same point from the practitioner side. Many open-weight releases publish only the trained weights; the training data and the full training code often stay private. Some open-weight licences are standard open-source licences, but many are custom agreements written by the model’s developer. So the availability of downloadable weights tells you nothing about whether your commercial use is permitted.
The licence shapes you will meet
In practice, open-weight licences fall into a few recognizable shapes. Sorting a candidate model into one of them is the fastest way to gauge how much review work it will take.
- Standard permissive open-source licences. Apache 2.0 and MIT are the familiar examples. They grant broad commercial rights with light conditions, typically attribution and notice retention. When an open-weight release uses one of these, the review is short.
- Standard reciprocal licences. These grant commercial use but attach conditions to derivative works. Whether those conditions reach your application, your fine-tune, or only the model itself is a question for a lawyer, not a README.
- Bespoke developer licences. Rodan notes that many open-weight releases use custom agreements written by the model’s developer. These are where commercial-use restrictions, usage thresholds, acceptable-use policies, and naming or branding conditions usually appear. They also tend to change between model versions.
Two related documents are worth keeping separate. A model’s licence governs the weights and any associated code. A model’s acceptable use policy is a separate document that can prohibit categories of use even where the licence grants broad rights. Read both.
A clause-by-clause checklist
These are the terms that most often decide whether a commercial deployment is straightforward, conditional, or blocked. Work through them in order and write down the answer for each.
1. Grant of commercial rights. Does the licence explicitly permit commercial use, or is it silent? Silence is not permission. Look for the words that grant rights rather than the words that describe the model.
2. Redistribution. If you ship the weights, a fine-tune, or a derivative model to customers, partners, or the public, you need a redistribution right. Internal use and redistribution are different permissions, and a licence can grant one without the other.
3. Attribution and notice. Many licences require you to carry a notice, name the model, or reproduce the licence text. In a product, that means deciding where the notice lives, whether in documentation, an about screen, or a NOTICE file, before release rather than after.
4. Usage thresholds. Some bespoke licences are free below a threshold and require a separate agreement above it. Thresholds are often expressed in monthly active users or revenue. Know which metric applies, what your current number is, and who monitors it as you grow.
5. Derivative works and naming. Fine-tuning, distillation, and model merging create derivative models. Check whether the licence permits that, whether the resulting model must carry the same licence, and whether you may use the original model’s name in your product or marketing.
6. Outputs and intellectual property. Establish who owns the model’s outputs and whether the licence makes any claim over them. Cooley lists intellectual property among the key legal risk areas for open-weight deployment, alongside licensing, privacy, security, and an evolving regulatory landscape.
7. Acceptable use and prohibited purposes. Even a permissive licence can sit alongside a policy that bans specific applications. If your product touches regulated advice, medical, financial, or safety-critical use, read this section closely.
8. Data, privacy, and training provenance. Self-hosting keeps prompts and documents inside infrastructure you control, which is a genuine advantage for regulated organisations. It does not answer questions about what data the model was trained on or what your obligations are when you process personal data through it.
9. Termination and change. Ask what happens if the licensor revokes or changes terms in a later version, and whether your existing deployment is grandfathered. Record the exact licence version you relied on.
10. Flow-down obligations. If you distribute anything built on the model, some obligations may pass to your users. Identify which ones and reflect them in your own terms.
A decision procedure you can repeat
- Classify the deployment. Internal-only use, a hosted product feature, or redistribution of weights and derivatives. Each has a different permission set.
- Collect the documents. The model card, the licence text, the acceptable use policy, and any commercial terms page, at the version you intend to deploy.
- Sort the licence into a shape. Standard permissive, standard reciprocal, or bespoke. This tells you how deep the review needs to go.
- Run the checklist above and record each answer with a reference to the exact clause.
- Escalate only the open questions to legal, with the deployment classification attached. A short, specific question gets a faster answer than asking whether the team can use the model at all.
- Record the decision centrally. Rodan recommends storing the result where the governance team can find it. Include the model version, the licence version, the date, and the deployment it covers.
- Set a review trigger. Re-check when the model is updated, when the licence changes, when your usage crosses a threshold, or when your deployment model changes.
An illustrative scenario
Consider a mid-sized company that wants to classify and summarise inbound support tickets with a self-hosted open-weight model, then later decides to ship the same capability inside its own product. The internal phase is usually the easy case: the model runs on the company’s own hardware, prompts and documents stay inside its infrastructure, and the main obligations are attribution and acceptable-use compliance. The product phase changes the analysis. If the capability is exposed to customers and the company redistributes a fine-tune, redistribution and flow-down clauses move to the centre, and any usage threshold now has to be monitored against a growing user base. This scenario is illustrative, not a reported case; the point is that the same model can move from low-risk to high-obligation without the licence changing at all.
Where self-hosting changes the calculus
Licensing is not the only reason to choose an open-weight model, and the technical case has strengthened. Rodan reports that in April 2026, Artificial Analysis found the leading open-weight models were six points behind the leading proprietary model on its Intelligence Index, while the gap remained wide on the hardest reasoning and agentic coding tests. For routine work such as summarising, classifying, and extracting, that may be close enough, and your own evaluation should decide it. Self-hosting also helps with data residency, version control, and predictable running costs, which is why regulated organisations consider it seriously.
None of those benefits remove licence obligations. Running a model entirely on your own hardware changes where the data goes, not what the licence permits.
When the model sits inside something you ship
If the model powers a WordPress plugin, an automation, or any component you distribute, the licence review has to cover the distributed artefact, not just your internal use. Attribution notices may need to appear in a plugin readme or an about page. Redistribution rights may be required for any bundled weights or adapters. If your automation sends customer content to a self-hosted endpoint, your own privacy documentation should describe that flow accurately. Treat the model as a dependency with a licence, in the same way you would treat a library, and pin the version you vetted.
What this guide cannot settle
Licence terms are drafted by the model’s developer and can change between releases; the sources here are a law-firm alert and a practitioner playbook, not a survey of every licence in circulation. Neither substitutes for advice on your specific facts. Where a term is ambiguous, or where your deployment is unusual, for example embedded in a regulated product, redistributed to third parties, or used at a scale that may cross a threshold, the correct next step is a scoped legal question rather than an engineering workaround.
Next steps
Pick the model you are actually considering, download its licence and acceptable use policy at the version you intend to run, and fill in the ten-item checklist. Record the result, note the version and date, and set a review trigger. If you have already worked through a licence that surprised you, whether a threshold you nearly missed or an attribution requirement that reached your product, share the term and the deployment context in the community so other teams can check the same clause.
FAQ
Can I use an open-weight model commercially?
Only if the licence grants that right. Some open-weight releases use standard open-source licences, but many use bespoke agreements written by the model's developer, so the licence text and any acceptable use policy decide the answer.
Is open weight the same as open source?
No. Open-weight releases typically publish trained weights only, while training data and full training code often stay private, and the licence may restrict commercial use, redistribution, or attribution.
Does self-hosting remove licence obligations?
No. Running the model on your own infrastructure can help with data residency, version control, and predictable running costs, but the licence still governs what you may do with the model and its outputs.
Do I need to re-check the licence after a model update?
Yes. Bespoke licences can change between model versions. Record the model version and licence version you vetted, and set a review trigger for updates and for changes in your own deployment.
Do I need a lawyer to review the licence?
Both sources present their material as guidance rather than legal advice and recommend that a qualified adviser confirm the final decision, particularly for redistribution, regulated use, or deployments near a usage threshold.