Rights and contracts
Does a data license cover future model versions?
By SourceX Editorial · Reviewed by Noah Loul ·
Short answer
A data license covers future model versions only if the grant says so. Read the model definition, the permitted-use clause and the term together: a grant tied to a named model usually stops there, while a grant to train models during the term often reaches successors. Decide the scope on purpose and treat broader rights as a priced concession.
Key takeaways
- Coverage of future model versions comes from the wording of the grant, not from industry custom.
- A defined 'Licensed Model' narrows a license; a plural grant to train 'models' during the term widens it.
- Training new versions during the term and retraining after the term ends are separate questions with separate clauses.
- Fine-tuned, distilled and merged derivatives should be named in or out, because silence invites a dispute.
- Broader version rights are a commercial concession and should be priced, approved and recorded as one.
What decides whether a license reaches the next model?#
Three clauses decide whether a data license reaches the next model: the definitions, the grant and the term. The definitions say what a 'model' or 'Licensed Model' is, the grant lists the verbs the buyer may perform with the records, and the term and termination clauses say when those verbs stop.
General counsel often read only the grant sentence and miss a definition buried pages earlier. A grant that says 'train the Licensed Model' means little until you know whether that defined term names one release, a product line, or 'any model developed by Licensee or its Affiliates'.
Where the agreement is silent, the answer turns on contract interpretation under the governing law, and both sides carry risk. Fixing the scope at signing is cheaper than arguing about it after a buyer announces a new model.
- Definitions: Licensed Model, Model, Derivative Model, Affiliate, Licensed Data.
- Grant: the verbs (train, fine-tune, evaluate, test, improve) and any purpose limit.
- Restrictions: release of weights, resale of data, use for third-party models.
- Term and effect of termination: what continues, what must stop, what must be deleted.
- Exclusivity and follow-on rights: whether scope over future models carries extra restrictions on you.
What do retraining, successor and version terms mean?#
Retraining, successor and version terms each answer a different question about time and scope. Mixing them up is the most common reason two parties sign the same words and walk away with different expectations.
| Term | What it usually covers | Question to settle |
|---|---|---|
| Retraining | Training an existing model again on the licensed records, often after a refresh delivery | Is retraining allowed after the term ends, or only during it? |
| Successor model | A later model in the same line that replaces the licensed one | Does the grant follow the product line or stop at the named release? |
| Version limit | Use tied to named releases or a defined model family | How is a new version identified, and who decides? |
| Derivative model | Fine-tuned, distilled, merged or compressed models built from a licensed model | Are derivatives covered, and may they be shipped to the buyer's customers? |
| Evaluation use | Testing or benchmarking a model against the records without training on them | Is evaluation allowed for models outside the licensed scope? |
Three illustrative clause options for version scope#
Version scope usually lands on one of three positions, from narrow to broad. The paraphrased options below show the shape of each; actual wording should come from counsel who knows the deal and the governing law.
Option one ties the grant to a named model: the buyer may train and evaluate the Licensed Model, defined as a specific model line and release. Option two follows a model family during the term: the buyer may train the Licensed Model and its successor versions released before the term ends. Option three covers any internal model: the buyer may train any model it develops, subject to firm limits on disclosing the records or releasing weights.
| Option | Supplier position | Buyer position | Typical guardrails |
|---|---|---|---|
| Named model only | Tightest control; a new deal is needed for the next model | Least flexible; may need to renegotiate at each release | Clear release identifier, evaluation-only use for other models |
| Model family during the term | Middle ground; scope ends with the term | Can carry records into successors without renegotiation | Notice of new model lines, no post-term retraining |
| Any internal model | Broadest grant; control rests on restrictions | Maximum flexibility across teams and products | No open-weight release, no resale of records, deletion at term end |
Is training during the term different from use after it?#
Training during the term and using a model after the term are different questions, and many disagreements come from treating them as one. A model trained while the license was active has already learned from the records, and many agreements let that model keep operating after the term ends.
The practical lever for a supplier is what happens to the records themselves. A supplier-friendly agreement usually bars any new training on the licensed records after the term, requires deletion of retained copies and asks for a written deletion certificate.
Successor rights interact with this. If the grant follows a model family but training must stop at term end, a successor released after that date cannot be trained on your records unless the license is renewed.
Illustrative: a vertical software company narrows the scope#
Illustrative: a fictional maker of scheduling software for commercial cleaning firms holds years of Zendesk tickets linked to Jira issues and GitHub pull requests. A model developer's draft agreement grants rights to 'develop, train and improve any current or future models' with no end to training rights.
The company's outside counsel proposes the model-family option instead: training on the current model line and successors released during the term, evaluation-only use for other models, no release of weights trained mainly on the records, and no retraining after the term. The buyer accepts the family scope but asks for a renewal option on refreshed tickets.
The signed agreement records the family scope, an end to training at term end, a deletion certificate and a renewal option the company may decline. The CEO and general counsel can explain in one sentence what the buyer may and may not do with the next model.
What to settle before agreeing to future versions#
Agreeing to future versions is safest when each related clause is settled in the same negotiation. The checklist below keeps the scope question from leaking into other clauses later.
- Define the model, the model family and how a new version is identified.
- State whether models released with open weights are in or out of scope.
- Say whether affiliates' models are covered or only the contracting entity's.
- Decide whether the buyer may build customer-specific fine-tunes from the records.
- Separate training rights during the term from any rights after it.
- Require deletion of retained copies and a signed deletion certificate at term end.
- Check how follow-on rights such as a renewal option or first refusal affect your future buyers.
- Link broader scope to price, because the buyer is receiving more.
How SourceX handles model version scope#
SourceX treats model version scope as a Rights question in the SourceX five-step transaction: Supply, Rights, Preparation, Approval and Delivery. The supplier decides how broad the grant may be, and nothing is delivered until the supplier approves the final terms.
The agreed model definition, any successor rights and the end-of-term training rule are recorded as permitted use in the SourceX Evidence Packet. When a buyer announces its next model, both sides can check that one entry instead of re-reading the whole agreement.
Frequently asked questions
Is a license to 'improve our services' broad enough to train new models?
It can be read that way, which is the problem. Purpose language such as improving services or products is often broad enough to cover new models, so a supplier who wants a narrower result should tie the grant to defined models and say plainly that other models are excluded except for evaluation.
Can a supplier ask for more when the buyer wants successor-model rights?
Commercially, yes. Broader scope, longer terms and exclusivity are concessions, and parties usually negotiate them together. There is no standard price for any of them; value is known only once a buyer engages with the specific records and terms.
What about models the buyer builds for its own customers?
Customer-specific models are a separate scope question. Some suppliers allow them only for the buyer's internal use, others allow them with a bar on handing over the records themselves. State the position, because a grant to train 'any model' may otherwise include models delivered to third parties.
Do open-weight model releases change the analysis?
Yes. Once weights are public, neither party can control downstream use or deletion. Suppliers often address open release directly, either excluding models trained mainly on their records or requiring consent before any public release of weights.
If the license is silent on future versions, who is right?
There is no general answer. A court or arbitrator would interpret the whole agreement under its governing law, including definitions and purpose language. Because both sides face uncertainty, it is better to settle the question in an amendment than to wait for a dispute.
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.