Can a buyer release open-weight models trained on your data?

When may a buyer publish model weights?#

A buyer may publish open-weight models trained on your data only if your license permits it, either expressly or because the permitted-use clause is broad enough to cover distribution. Open-weight means the trained model's parameters are published for anyone to download and run, as opposed to keeping the model behind an API or inside the company.

Many data licenses focus on what the buyer may do with the dataset and say little about the models. That gap matters because a model is a separate artifact: the buyer usually owns it, and unless the license restricts distribution, the buyer may argue it can do what it likes with its own model.

For a CEO or founder, the practical point is simple. Decide your position on public release before the term sheet, not after a draft model card appears.

Why public weights change your risk#

Public weights change your risk because control passes from one licensee to anyone who downloads the model. With a hosted model, the buyer can filter outputs, monitor use and retrain if a problem appears. With released weights, none of those controls follow the copies.

The concrete concerns are memorization and extraction, where a model reproduces fragments of training records when prompted; fine-tuning by third parties that strips safety layers; and competition, where a freely available model trained on your support or engineering history helps others offer services like yours.

Preparation lowers the first risk. Removing names, contact details, credentials and client identifiers before delivery means whatever a model might reproduce is less sensitive. It does nothing for the competitive question, which is commercial and belongs in the contract.

Release modes compared#

Release modes range from internal use to fully public weights, and each calls for a different license term. Most suppliers are comfortable with the first two rows, which match how many commercial models are offered; the last two deserve explicit negotiation.

Clauses that restrict or permit release#

Clauses that restrict or permit release work best when they name both the artifacts and the acts. The outline below covers the usual options; counsel adapts the wording to the deal.

The derived data clause is the one most often missed. A buyer could train a closed model on your records, use it to generate synthetic examples, then release an open model trained on those examples. Without that coverage, a distribution clause can be sidestepped.

• 1Model distribution restriction: the buyer may not publish, distribute or make available the weights of any model trained, fine-tuned or evaluated on the licensed data, except as permitted.
• 2Derived data coverage: the restriction also applies to models trained on synthetic data, labels or distilled outputs generated from the licensed data.
• 3Conditional permission: release is allowed only after written approval, memorization testing by an agreed method and exclusion of specified record types.
• 4Training-mix threshold: release may be allowed where the licensed data is a minor component of training, with the threshold defined in the agreement.
• 5Notice and naming: whether the buyer must notify you before release, and whether your company may or may not be named.
• 6Survival: restrictions on released or releasable models survive termination of the license.

Choosing your position on public release#

Your position on public release should follow from what the records reveal and how much the licensed data would shape the released model. A supplier licensing prepared engineering issues has different exposure from one licensing customer conversations that reflect its service playbook.

Ask the buyer early whether it publishes models, at what stage of development, and under which licenses. A developer that never releases weights may accept a flat prohibition easily; one with a research release program will want a defined path, and that path is better negotiated in the term sheet than in the final draft.

What happens if the license ends after a release?#

Ending a license after an open-weight release cannot undo the release. Deletion obligations bind the buyer, not the people who already downloaded the weights, so the remaining remedies are damages, indemnities and restrictions on the buyer's future models.

This is why release is treated differently from other permitted uses. A license can usually require deletion of the dataset and, in some negotiations, retirement or retraining of affected models on termination. Public weights take those options off the table, so a decision to allow release is effectively permanent.

Copyright law may not fill the gap. In Getty Images v Stability AI, decided on November 4, 2025, the High Court of England and Wales held that the Stable Diffusion model weights did not store or reproduce Getty's works and so were not an infringing copy; Getty was later granted permission to appeal. If courts treat weights that way, rights in your records may not follow a released model, which leaves the license as your main control.

Illustrative: a helpdesk software company negotiates model rights#

Illustrative: a fictional helpdesk software company prepares a package of its own Zendesk tickets linked to Jira issues and GitHub fixes. A model developer proposes a license that allows training and commercial deployment but says nothing about distribution.

The CEO asks counsel to add a model distribution restriction covering open-weight release and models trained on synthetic data derived from the package. The developer explains that it publishes some smaller research models. The parties agree that public release requires the company's written approval, a memorization test against the delivered records and exclusion of customer-facing reply text, with the restriction surviving termination.

How SourceX treats model distribution terms#

SourceX records permitted use, including any model distribution terms, in the SourceX Evidence Packet for each package, alongside provenance, licensing rights, the privacy record and release authorization. The supplier approves those terms in the Approval step of the SourceX five-step transaction, before Delivery.

SourceX's dataset rights, including any training use, are set out in the signed supplier agreement, and data is licensed, not sold outright, so the supplier keeps ownership of its records and decides how far model rights extend.