Back to Blog

    November 18, 2026

    AI Model License Agreement: How to License an AI Model (Key Clauses and Pricing)

    Drafting a deal? Use the free SaaS licensing agreement template.

    Looking for SaaS to license? Browse live listings or explore products by category. New to the deal structures? Compare licence and white-label types or check a term in the glossary. Explore the white-label SaaS marketplace, software to license or license vs build. SaaS founder? List your Product for licensing.

    Quick answer: An AI model license agreement gives a buyer the right to use a machine learning model under set conditions while the developer keeps ownership. It must say how the model is delivered (API access or model weights), what it may be used for, whether the buyer may fine-tune it and who owns the fine-tuned version, who owns outputs, whether customer data may be used to improve the model, and who is responsible if the model or its outputs infringe someone's rights. If your model is built on an open base model, its license terms flow into yours.

    More founders now have a model worth licensing: a fine-tuned language model for a niche, a vision model trained on proprietary images, a forecasting model built on years of domain data. Standard software terms don't cover the questions a model raises, such as who owns a fine-tuned copy or what happens to outputs after the deal ends. This guide explains the AI model license types, the clauses that matter most, how to price access, and the checks both sides should run.

    Key takeaways

    • Decide first how the model is delivered: hosted API access and a license to the model weights need very different terms.
    • Fine-tuning, derived models and outputs each need a stated owner; silence causes the biggest disputes.
    • If your model is a fine-tune of an open model, the base model's license (attribution, use limits, user thresholds) still applies.
    • Say whether customer prompts and data may be used to improve the model; buyers increasingly ask for a clear no.
    • Price follows rights: weights, exclusivity and resale cost more than metered API access for internal use.

    What an AI model license is

    An AI model license is the set of permissions a model's developer (the licensor) gives a user (the licensee). The developer keeps ownership of the model; the licensee gets defined rights to use it. The agreement covers four things standard software licenses don't handle well: the model weights, changes made by training it further, the outputs it produces, and the data that flows through it.

    Licensing a model is close to licensing software or an API, and many of the same tools apply: license types such as exclusive or non-exclusive, perpetual or subscription, and the delivery terms in an API license agreement. What makes an AI model license different is that the asset learns. Training it on new data creates something new, and the questions of who owns that and what happens to it at the end of the deal must be answered in writing.

    AI model license types compared

    License typeWhat the licensee getsTypical termsBest for
    Permissive open license (such as Apache 2.0 or MIT)Weights and code to use, change and redistributeKeep notices; few or no use limitsDevelopers who want wide adoption
    Open license with use restrictionsWeights to use and change, within stated limitsBanned uses that pass on to downstream usersOpen models with responsible-use rules
    Custom community licenseWeights for free use up to set limitsAttribution, user or revenue thresholds, separate deal above themLarge model publishers
    Commercial API licenseAccess to a hosted model, no weightsUsage pricing, acceptable use policy, data termsMost paid models and model-powered products
    Commercial weights or on-premise licenseA copy of the weights to run on the buyer's own systemsSecurity duties, deployment limits, audit, higher feeRegulated buyers and enterprise deployments
    White-label or OEM model licenseThe right to offer the model inside the buyer's own branded productWholesale pricing, branding and support termsAgencies, resellers and software vendors

    API access vs a license to the weights

    The first decision is how the buyer gets the model, because it changes almost every clause.

    • Hosted API access: the developer runs the model and the buyer sends requests. The developer keeps control of the weights, can update the model and can enforce limits through the API. The agreement focuses on usage limits, uptime, data handling and price per use.
    • Weights license: the buyer receives a copy of the model to run on their own systems. This suits buyers with strict data rules or high volumes, but once weights leave your servers you rely on the contract to control them. The agreement needs security duties, deployment limits (which systems, how many instances), no reverse engineering or extraction, audit rights and return or deletion on termination.

    Many developers offer both: API access by default, and a weights license for larger customers at a higher price.

    Open, community and commercial model licenses

    Publicly released models come with very different licenses, and the label "open" covers several of them.

    • Permissive licenses such as Apache 2.0 and MIT allow use, change and redistribution, including commercially, as long as notices are kept. Apache 2.0 also includes an express patent license.
    • Licenses with use restrictions, such as the responsible-AI (RAIL) family, allow broad use but list banned uses, and require anyone who passes the model on to include the same restrictions.
    • Custom community licenses from large model publishers allow free use within limits, which can include attribution rules, an acceptable use policy and a requirement to get a separate license above a large user threshold.
    • Commercial licenses are negotiated or offered on standard terms of service and are what most paid models use.

    If your model is a fine-tune of an open base model, its license travels with you. Before you license your model to anyone, check that the base license allows commercial use and redistribution of a derived model, what notices and naming rules apply, and whether its banned-use list must be passed on to your customers.

    AI model license agreement: 12 clauses that matter

    1. Grant and delivery

    Who may use the model, how (API or weights), for how long, and whether the license is exclusive in any field, industry or territory. See exclusive vs non-exclusive licenses for how to limit exclusivity.

    2. Permitted and prohibited uses

    An acceptable use policy listing banned uses, plus any high-risk uses the buyer must not attempt without agreement, such as decisions about credit, employment or medical care.

    3. Fine-tuning and derived models

    Whether the buyer may train the model further, who owns the result, and whether the buyer may keep using a fine-tuned version after the license ends.

    4. Outputs

    Who owns outputs, whether the developer may use them, and any limits, for example a ban on using outputs to train a competing model.

    5. Customer data and model improvement

    Whether prompts, files and outputs may be used to train or improve the model. Many business buyers require a clear no, or an opt-in.

    6. Training data and provenance

    What the developer promises about the data the model was trained on: that it had the right to use it and that the training complied with the licenses that applied. A data licensing agreement is where those rights are usually secured.

    7. Performance and changes

    Models are probabilistic, so developers rarely promise accuracy. Agree what is promised instead (availability, response times, a named model version) and how much notice the buyer gets before a model is changed or retired.

    8. Intellectual property indemnity

    Who pays if a third party claims the model or its outputs infringe their rights. Developers often limit this to unmodified use of the model within the acceptable use policy.

    9. Liability caps

    The ceiling on what each side can claim, usually tied to fees paid, with any exceptions spelled out.

    10. Security and confidentiality

    For weights licenses especially: where the model may run, who may access it, how it is stored, and a duty to report leaks.

    11. Compliance with AI rules

    AI regulation, such as the EU AI Act, places duties on both the developers and the deployers of some AI systems. Agree who supplies documentation, who handles transparency to end users and who carries which duty.

    12. Termination and what survives

    What must stop, what must be deleted (weights, fine-tuned copies, cached outputs) and what the buyer may keep. This matters most for weights and fine-tuned models, which are hard to recall once deployed.

    This is general information, not legal advice. AI law is changing quickly, so have a lawyer review the agreement for your jurisdiction and use case.

    How to price an AI model license

    Price follows the rights granted and the cost of serving the model. Common structures:

    • Usage pricing: per token, call, image or prediction, often with monthly minimums. It fits API access and tracks your own compute costs.
    • Subscription tiers: a monthly fee with a usage allowance, simpler for small buyers.
    • Annual weights fee: a yearly fee for an on-premise copy, sometimes per deployment or per server.
    • Per-customer or per-seat fees: common when a partner offers the model inside their own branded product.
    • Revenue share: a percentage of what the buyer earns from a product the model powers; see the revenue share agreement guide for reporting and audit terms.

    Our guide to AI agent pricing models covers per-usage, per-seat and outcome pricing in more depth, and most of it applies to models too.

    Checklist before you license a model

    For model developers:

    • Confirm you have the rights to the training data and that the base model license allows what you plan.
    • Choose API, weights or both, and set different terms for each.
    • Write the acceptable use policy and the rules for fine-tuning and outputs.
    • Set usage limits, a version and deprecation policy, and a price per tier.

    For buyers:

    • Test the model on your own data before you sign, and agree which version you are buying.
    • Get a clear answer on whether your data trains the model.
    • Check output ownership, IP indemnity and its exclusions.
    • Plan for the model being changed or retired, and for what you keep if the deal ends.

    Where LicenseSaaS fits

    LicenseSaaS is a marketplace where founders list software, AI tools, AI agents, APIs and datasets for licensing and white-labeling. If you have built a model or a model-powered product, list it for free and check buyer requests for what people are asking for. If you need one, browse what's available to license, look at white-label AI software, and read how to license AI agents if the model will run inside an agent. To sell the model or its training data, see how to sell AI agents and how to sell data to AI companies. LicenseSaaS does not verify listings, so test the model and check its terms yourself before you sign.

    Frequently asked questions

    What is an AI model license?

    It is the permission a model's developer gives others to use the model, and the rules that come with it: how it is accessed, what it may be used for, whether it may be changed or redistributed, and who owns what it produces. Open models publish a standard license; commercial models use a negotiated agreement or terms of service.

    Can I use an open-source AI model commercially?

    Often yes, but read the license. Permissive licenses such as Apache 2.0 and MIT allow commercial use with notice requirements. Some model licenses add banned uses, attribution rules or limits above a large number of users, and some data or model licenses exclude commercial use entirely.

    Who owns a fine-tuned version of a licensed model?

    Whoever the agreement says. Commonly the base model's developer keeps the base model, and the licensee owns its own training data and the changes it made, with the right to use the fine-tuned model only while the license lasts. Agree this before fine-tuning starts.

    Who owns the outputs of a licensed AI model?

    Most commercial agreements give the customer the outputs, subject to the acceptable use policy, but terms vary and the law on protecting AI-generated material is still developing. State output ownership in the agreement, and whether the developer may use outputs or prompts to improve the model.

    How do you price an AI model license?

    Common models are usage pricing by tokens, calls or predictions for API access; an annual fee for on-premise weights; per-seat or per-customer fees in white-label deals; and revenue share when the model powers the buyer's product. Exclusivity and wider rights cost more.

    Take the next step

    Discover SaaS products to license, browse by category, or list your own product on LicenseSaaS.