Licensing a model is not licensing software. Software has a defined boundary: the code, the documentation, the right to run it. A model has weights that can be copied, outputs that can be used to train a competitor, and behaviour that changes with fine-tuning. Each of those is a leakage route that a conventional software licence does not address.

When the licensee is in China, three further layers apply: the trade secret regime that protects the weights, the data rules that govern what the model processes, and the regulatory regime for AI services offered to the Chinese public. A licence drafted without those layers reads well and fails at the point of enforcement, which is usually the point at which the licensor has already lost the asset.

This note deals with the parts of a model licence that decide whether the licensor still controls the model in year three: how the weights are defined, how they are delivered, how they are escrowed, how they are audited, and how derivative models are treated. The grant, the data schedule and the exit are covered as they bear on those five questions.

What is actually being licensed

A model is a bundle, and each element has a different legal character.

ElementTypical legal characterDefault risk
Model weights and parametersTrade secretCopyable, hard to detect
Inference code and serving stackCopyright, sometimes patentReverse engineering
Training pipeline and recipesTrade secretDisclosure reveals how to rebuild
Training dataTrade secret, database rights, PIPLRegulatory exposure and re-identification
Documentation and promptsCopyrightLow commercial value, high leakage value
Trade marks and model nameTrade markBrand dilution and false affiliation

The agreement should schedule these separately. A single defined term, “the Model”, that covers all six makes it impossible to grant different rights over different elements, and it makes the termination and deletion provisions unworkable, because deletion of “the Model” has no defined meaning when the Model includes a trade mark.

Defining the weights precisely

A clause that licenses “the model weights” has licensed a phrase. The weights are a set of numerical parameters that only become a working model when combined with several other artefacts, and the definition should identify the set rather than the concept.

Six components are usually needed for the weights to be usable: the parameter tensors themselves, in a stated precision and format; the architecture configuration, which defines how the tensors are arranged; the tokeniser and its vocabulary; the checkpoint identity, meaning the training step or release tag from which the parameters were taken; any adapters or additional heads shipped with the checkpoint; and the inference code and serving configuration that turn a tensor file into a service.

Two further elements are frequently overlooked and matter commercially. Safety classifiers and content filters may be separate models with separate weights, and a licence that covers “the model” may or may not cover them. System prompts and evaluation harnesses are not weights at all, but they are often the most valuable operational knowledge in the package, and a licensee that receives them receives a substantial part of the licensor’s engineering.

Precision deserves a sentence in the schedule. An FP32 checkpoint and an eight-bit quantised checkpoint are different assets with different performance, and a licensee that receives the full-precision weights can produce the quantised version itself. Where the licensor intends to ship only a quantised build, the schedule should say so, and the delivery provisions should make the full-precision weights a separate, controlled item.

The schedule as a control instrument

The schedule should read like a manifest, not a description. For each item: a name, a version or release tag, a hash, a size, a delivery medium, and a statement of what the licensee may do with it. That format does three things at once. It makes delivery verifiable, because the hash can be checked on arrival. It makes the audit meaningful, because the deployed artefact can be compared with the delivered one. And it makes deletion testable, because a certificate that names six items can be checked against a register that names the same six.

Delivery: API, on-premise and edge

The delivery model determines the risk, and it is a commercial decision that should be recorded in the agreement rather than discovered at deployment.

API access keeps the weights on the licensor’s infrastructure. The licensee receives outputs, not parameters. Copying is limited to what can be inferred from outputs, and access can be revoked. The trade-off is dependence on the licensor’s availability, latency and pricing, which matters for industrial deployments where a stopped inference service stops a production line.

On-premise deployment places the weights in the licensee’s environment. It is often required for data residency, latency or security reasons, particularly for robotics and industrial AI. It also places the asset where copying is technically trivial. On-premise licences need stronger controls: encryption at rest with key management the licensor can verify, deployment attestation, audit rights, watermarking and a higher royalty.

Edge deployment on devices is the hardest to control. Once weights are on a shipped device, extraction is a matter of effort rather than permission. The commercial answer is often to limit what is shipped, such as a distilled model, a subset of capabilities or a secure element, rather than to rely on contract alone.

The delivery schedule and chain of custody

Delivery is an event with a record, and the agreement should treat it as one.

A workable delivery clause specifies the medium, the encryption, the key channel and the acknowledgement. The medium might be an encrypted archive over a secure channel, a signed container image from a registry the licensor controls, or physical media for an air-gapped site. The key should travel by a different channel from the archive, so that a single interception does not yield both. The acknowledgement should be signed by a named individual at the licensee, should list the hashes received, and should state where the artefacts were installed.

Chain of custody is the part that foreign licensors skip and later regret. The licensee should maintain a register of every copy of the weights: the environment, the host identifiers, the date of installation, the person responsible, and the date of removal. The register should be available on audit and should be updated whenever the deployment changes. Without it, a deletion certificate is a statement of intent rather than a statement of fact, and an audit cannot establish whether a copy exists outside the boundary.

For air-gapped sites, the physical route needs the same discipline. A courier record, a sealed package, an installation supervised by the licensor or by a nominated engineer, and a signed installation report are the minimum. A licensor that ships a drive and receives a thank-you note has no evidence of what was installed or where.

Acceptance, and why it is not delivery

Delivery and acceptance are different events, and the distinction has commercial consequences. Delivery transfers possession; acceptance triggers payment and starts the warranty period. A clause that treats delivery as acceptance means the licensee pays for a package it has not run.

Acceptance should be defined by test rather than by impression. A benchmark suite, a set of reference inputs with expected outputs within a tolerance, a latency and throughput target on the licensee’s hardware, and a period within which the licensee must either accept or report a failure. The period matters: an open-ended acceptance right is a payment delay.

Model performance warranties are difficult to draft and easy to over-promise. A warranty that the model “will perform in accordance with the documentation” is enforceable; a warranty that it “will meet the licensee’s requirements” is not, because the requirements are not in the contract. Where the licensee insists on an accuracy commitment, tie it to a named evaluation set delivered with the model, and state what happens when the licensee’s own data performs worse than the evaluation set. It usually will.

Protecting the weights as trade secrets

Chinese trade secret protection requires three elements: the information is not public, it has commercial value, and the holder has taken corresponding confidentiality measures. For weights, the third element is the operative one, and it is the one a licensor controls entirely.

Measures that survive scrutiny

Measures recognised in practice include access controls on model repositories, named-user download logs, encryption at rest, individual confidentiality undertakings from the engineers with access, restrictions on local copies, and a disclosure record for each transfer.

The measures need to be contemporaneous with the disclosure, not assembled after a dispute. A licensor that signs NDAs in year one, gives repository access to a shared account in year two, and produces a security policy in year four has a policy, not a measure. Tribunals and courts look for the practice that actually governed access at the time the information was disclosed.

For an on-premise deployment, the measures should extend into the licensee’s environment: a defined deployment boundary, a prohibition on copying outside it, technical attestation where feasible, and an audit right. A licence that permits on-premise deployment but says nothing about copying is relying on an assumption the technology does not support.

Marking, segregation and a written exit procedure strengthen the position at modest cost. Each artefact should carry a confidentiality legend and a recipient identifier. No single engineer should hold the complete package. And access should be revoked on the day a person leaves rather than at the end of the month.

What defeats the claim

Four failures recur in Chinese disputes.

The weights were downloadable from a public repository at some point, and the licensor’s own earlier release defeats secrecy. A model that was briefly open-weight and later closed is a difficult trade secret case, and the licence should not assume otherwise.

The licensor cannot identify the secret with precision. “Our model” is not a secret; “the parameters in release 4.2, identified by hash, together with the training recipe in Annex C” is. A trade secret claim needs a defined subject matter, and a vague claim invites the court to conclude that the licensor is protecting a general skill rather than specific information.

The licensee can show it developed the same capability independently. Independent development is a defence, and the more capable the licensee’s own research team, the more plausible the defence. That is one reason the audit trail matters: it is the licensor’s evidence that the licensee had access.

The measures were contractual only. An NDA with no access control, no marking and no log is weak evidence that the licensor treated the information as secret, even though it is strong evidence of an obligation. The two are different questions and the licensor needs both.

Restrictions that carry the commercial weight

Three restrictions do most of the work in a model licence, and each needs to be named rather than inferred.

No distillation or competing training. A general reverse-engineering prohibition may not obviously cover the training of a competing model on the licensed model’s outputs. Name distillation, and prohibit the use of outputs to train or improve a competing model. This is the leakage route that matters most for foundation models, and it is the one most often left unaddressed.

No sublicensing or redistribution. Including to affiliates, contractors and customers, unless named. A model embedded in a licensee’s product and shipped to customers is a redistribution, and the agreement should say whether that is permitted, on what royalty and with what restrictions on extraction.

No removal of safeguards. Where the model carries safety filters, watermarking or usage restrictions, the licence should prohibit their removal and treat removal as a material breach.

Naming derivative uses

A restriction works when it names the act. The list below is the one a technical team will recognise, and each item should appear in the prohibited-uses clause or be expressly permitted.

ActWhat it producesWhy the licensor cares
Full fine-tuningA new checkpoint derived from the base weightsThe licensee holds a near-copy of the asset
Parameter-efficient tuning (adapters)A small set of weights that modifies behaviourCheap to produce, hard to detect, easy to merge
Continued pre-trainingA model with the base as its starting pointClosest thing to a rebuild
Distillation from outputsA smaller model trained on generated dataNo weight copying, no technical trace
Quantisation and pruningA compressed deployment buildLegitimate operationally, and a route to redistribution
Merging and model surgeryA combined or modified modelObscures provenance and may remove safeguards
Reinforcement learning on outputsA model tuned on preference dataImproves a competitor using the licensor’s capability

A clause that prohibits “reverse engineering” and nothing else leaves most of that table unaddressed. A clause that prohibits the table and nothing else leaves the licensee unable to deploy, because quantisation is a normal production step. The clause has to distinguish the acts that are permitted for deployment from the acts that create a new asset.

Redistribution through products

Where the model is embedded in a product, the customer receives a copy of the weights in some form, even if the form is a compiled binary on a device. The licence should say what the customer may do: use the product, not extract the model; not run the model outside the product; not use the product’s outputs to train a competing model. Those obligations have to reach the customer, which means either a click-through term in the product or a contractual flow-down through the licensee’s distribution chain.

Where the licensee sells through distributors, the flow-down obligation should name the distributor agreement as the vehicle and require the licensee to enforce it. A licensor that has no visibility of the downstream terms is relying on a chain of contracts it has never seen.

Derivative models

Fine-tuning creates a derivative that is partly the licensor’s and partly the licensee’s. The licence has to allocate it before the work starts, because after the work starts the licensee has a model it is using in production and the negotiation is no longer about drafting.

A working taxonomy

Derivatives are not a single category, and treating them as one produces clauses that are either unenforceable or unworkable.

A full fine-tune updates all parameters on new data. The result is a checkpoint that is structurally identical to the base and behaviourally different, and the licensee holds something close to a copy of the asset. An adapter, by contrast, is a small set of additional parameters trained while the base is frozen. It is cheap, portable, and can be merged into the base later, which is why it is the most common form of customisation and the hardest to police.

Continued pre-training uses the base as the starting point for further training on a large corpus. It produces a model that is further from the original than a fine-tune and closer to a rebuild. Distillation trains a smaller model on the outputs of the larger one. It requires no access to the weights at all, which is why it is the route that contractual controls reach least well and the one that matters most for a licensor whose model is served by API.

Quantisation and pruning compress the model for deployment. They are operational necessities and, in a licence that prohibits modification, technical breaches that every licensee commits. Merging combines two models or an adapter and a base into one set of weights. Model surgery edits the weights directly, and is the usual method for removing safety behaviour.

The agreement should define “Derivative Model” by reference to this taxonomy, and should say for each category whether it is permitted, who owns it, and what happens to it on termination. A single sentence covering all seven will be read in whatever way favours the party that did not draft it.

Ownership by contribution

Three allocations are common: the licensor owns the fine-tuned weights and grants the licensee a licence back; the licensee owns them and grants the licensor a licence to use them; or ownership is split by contribution, with the base weights remaining the licensor’s and the adapter remaining the licensee’s.

The third is usually the most workable, because it reflects what each party contributed. It requires the agreement to define the boundary between the base model and the adapter, and to address what happens if the licensee merges the adapter into the base weights, which is technically common and legally ambiguous.

Two refinements make the split operable. First, define the boundary by artefact rather than by idea: the base checkpoint by hash, the adapter as the parameter set produced by a named training run, and the merged model as a new item whose ownership follows the base. Second, require the licensee to register each derivative with the licensor, with the training data category, the method and the resulting artefact hash. Registration is a light obligation that converts an invisible act into a record, and it is the difference between an audit that works and an audit that relies on confession.

Merging, quantisation and model surgery

Merging deserves its own clause because it changes the character of both inputs. Where an adapter is merged into a base, the merged model is a copy of the base with modifications, and the licensor’s interest in it is at least as strong as its interest in the adapter. The agreement should state that a merged model is a Derivative Model owned by the licensor, that the licensee’s rights in the adapter survive as a licence, and that the merged model is subject to the same restrictions as the base.

Quantisation should be permitted expressly, on conditions: the compressed build may be used only for the licensed deployment, may not be distributed separately from the product, and must be derived from a licensed copy. That converts a routine engineering step into a permitted act with a boundary, rather than a breach that the licensor will never enforce.

Model surgery that removes safety behaviour is a different matter. It should be prohibited, and the prohibition should extend to using a third-party tool to achieve the same result, because the act is the removal and not the method.

Derivatives of derivatives

The chain problem appears once a licensee’s derivative is itself licensed, sold or embedded in a product supplied to a third party. Each link inherits the restrictions of the link above it, and each link is a point at which the licensor loses visibility.

Three provisions control the chain. A prohibition on licensing or distributing a Derivative Model outside the licensee’s own products without consent. A flow-down requirement, so that any permitted distribution carries the same restrictions. And a termination provision that reaches the whole chain, so that on termination the licensee must procure deletion of its own copies and of copies held by its customers and contractors, to the extent it has the right to do so.

Where the licensee cannot procure deletion downstream, the agreement should say what happens instead: a sunset period, a continuing royalty for the installed base, and a prohibition on new deployments. The alternative is a deletion obligation the licensee cannot perform and will not attempt.

Escrow and continuity

Escrow is often proposed by Chinese licensees as a condition of an on-premise or hosted deal, and it is often accepted without examining what it is for.

Escrow is not a security measure. It does not prevent copying, and it does not protect the licensor’s trade secret, because the escrow agent holds a copy by design. What escrow does is protect the licensee against the licensor’s disappearance: insolvency, withdrawal from the market, or a failure to maintain the software on which a production line depends. That is a legitimate commercial concern for a licensee that has built a factory around a model, and it is a concern that a licensor can price.

What goes into the escrow

A deposit that contains only the weights is not usable. The licensee needs the artefacts required to run the model in the licensee’s environment, which usually means the weights at a stated precision, the inference code, the tokeniser, the configuration, the build and deployment instructions, the evaluation harness and the documentation. Without the build instructions, the deposit is a file; with them, it is a system.

The deposit should be versioned. Each release that the licensee is entitled to receive should trigger an update within a stated period, and the escrow agreement should record which versions are held. A deposit that stops being updated after the first year gives the licensee a model it stopped using two years ago.

Verification is the step that distinguishes a real escrow from a filing cabinet. The escrow agent, or an independent technical reviewer, should build and run the deposited artefacts in a clean environment and confirm that the result matches the delivered system’s reference outputs. Where verification is not feasible at deposit, it should be performed at release, because a release trigger that produces a non-functional deposit is worse than no escrow at all.

Release triggers

The triggers should be objective and narrow. Insolvency or administration of the licensor, cessation of the licensor’s business, and a sustained failure to provide support that the licensor is contractually obliged to provide are the usual three. A trigger that the licensee can engineer, such as any dispute over support quality, converts the escrow into an option over the asset.

Each trigger should have a procedure: notice to the licensor, a cure period where the trigger is a failure, a determination mechanism where the parties disagree, and release to a named beneficiary rather than to the licensee generally. Where the licensee is a group, the beneficiary should be the entity that operates the deployment, because that is the entity that needs continuity.

The licence that attaches on release

Release is not a sale. The escrow agreement should attach a licence to the released materials, and that licence should be narrower than the commercial licence in one respect and wider in another: narrower, in that it permits use only to support the licensee’s existing deployment and not to develop new products or new fields; wider, in that it survives the termination of the main agreement for as long as the deployment continues.

The released licence should also preserve confidentiality, prohibit sublicensing, and require a continuing royalty where the commercial licence provided for one. A release that gives the licensee a perpetual royalty-free right to the weights is not escrow; it is a sale at the price of a licence.

Audit and technical verification

An audit right over a model deployment is more intrusive than an audit right over a royalty report, and it should be drafted with that in mind.

The clause should define what may be inspected, such as deployment logs, access records, configuration reports and deletion certificates, and what may not, which usually includes the licensee’s customer data and proprietary code. It should set a notice period, a frequency limit and a confidentiality obligation on the auditor.

Verifying a checkpoint

The most useful audit is the cheapest one: a hash comparison. The licensee reports the hash of the deployed checkpoint, the licensor compares it with the delivery manifest, and the comparison establishes whether the deployed artefact is the delivered one. That takes minutes and answers the question that matters most.

Hash comparison does not answer everything. A licensee that has modified the weights will report a different hash, which proves modification but not the nature of it. A licensee that runs an unregistered copy will not report it at all. The audit therefore needs a second layer: the licence server’s token records, which show where inference was requested and from which deployment identifier; the telemetry the deployment emits, where the deployment is connected; and, where neither is available, an on-site inspection with read-only access to the deployment host.

Telemetry raises a data question that is easy to overlook. Logs from a Chinese deployment that are transmitted to a licensor outside China are a cross-border transfer of personal information if they contain user identifiers or query content. The data schedule should say what the telemetry contains, where it goes and on what legal basis, and the technical team should be told, because the default configuration of most observability tools sends more than the licensor needs.

Air-gapped deployments

Where the licensee operates in an environment with no outbound connectivity, the technical controls may be limited to a signed deployment package and an audit right. The agreement should reflect the actual environment rather than a standard template.

In an air-gapped site, the practical controls are a signed installation package, a local licence file bound to the host, a written installation report, and periodic on-site verification by the licensor or an independent assessor. Some licensors use a challenge-response mechanism in which the deployment generates a signed report that the licensee carries out on media. It is slower than telemetry and it works.

Where the licensee is a state-owned enterprise or a company with security-sensitive operations, the audit may be resisted on security grounds. The agreement should provide an alternative, such as a third-party certification or a report from an approved assessor.

What counts as evidence

Four categories of evidence are worth designing for at signature, because none of them can be created after the fact. The delivery record, which proves what was transferred and when. The deployment register, which proves where copies were installed. The technical trail, which proves where inference occurred. And the output watermark, which proves that a model found in the market derives from the licensor’s.

A licensor with all four can prove a case. A licensor with a signed contract and no records can prove an obligation and nothing else, which in a trade secret action is usually not enough.

How weights leak, and the controls that answer it

Weights rarely leave through the front door. They leave through the routes that the licence did not mention.

The first is the engineer’s laptop. A deployment that permits local inference invites a copy of the checkpoint onto a workstation, a personal drive or a cloud account. The copy is invisible to the licensor until a competing model appears with a familiar behavioural signature.

The second is the contract manufacturer. Where the model is embedded in a device, the firmware image may be extracted from the hardware, and the extraction may be performed by a factory that was never a party to the licence.

The third is the fine-tuning pipeline. A licensee that is permitted to fine-tune will hold a copy of the base weights in its training environment, and the environment is usually less controlled than the production deployment.

The fourth is the output channel. A licensee that is permitted to generate large volumes of outputs can use them to train a smaller model, which is cheaper than copying the weights and harder to detect.

A licence that addresses only copying, and not these four routes, addresses the least likely one.

Licence servers, attestation and technical controls

Contractual controls work better when they are backed by technical ones.

A licence server that issues short-lived tokens, bound to a deployment identifier, gives the licensor visibility of where inference is happening and the ability to revoke. Attestation, where the deployment environment reports its configuration to the licensor, detects a copy running outside the agreed boundary. Watermarking of outputs provides a way to identify the source of a model if a copy appears in the market.

None of these is perfect. All of them raise the cost of leakage and create evidence. The licence should describe the controls, require the licensee to maintain them, and treat their removal or circumvention as a material breach.

The controls should also be proportionate to the price. A licence server on a low-value pilot deployment costs more to operate than the royalty it protects, and a licensee that is asked to accept intrusive monitoring for a small engagement will price the intrusion into the deal. The control regime should scale with the value at risk and with the delivery model, which is one reason the delivery decision belongs in the commercial negotiation rather than in the implementation plan.

Open-weight models and the licence chain

A growing number of Chinese partners will propose to use an open-weight model rather than license a proprietary one. The commercial logic is obvious, and the legal analysis changes.

An open-weight licence is still a licence. It carries terms on commercial use, attribution, acceptable use and, in some cases, restrictions on training other models. A licensee that incorporates an open-weight model into a product must comply with those terms, and the supplier of the product inherits the obligation.

Where the licensor’s technology is combined with an open-weight model, the agreement should address the combination: which terms govern the combined work, whether the open-weight terms impose obligations that the licensor cannot meet, and who is responsible for compliance. A product that ships with an open-weight model and a proprietary model may carry two sets of obligations that conflict.

Where the obligations land

The consequences are practical, and each has a place in the agreement.

The terms travel with the weights. An open-weight licence that requires attribution, or that restricts commercial use above a stated scale, binds the party that distributes the resulting product, which may be the licensee and may be the licensee’s customer. The agreement should require the licensee to pass the terms on and to evidence that it has done so.

The terms change between versions. Open-weight licences are revised, and a model that was permissive at release 1 may carry additional conditions at release 4. The schedule should record the model, the version and the licence text, and the licensee should warrant that it has complied with the version it deployed.

The terms may conflict with the licensor’s own restrictions. A licence that requires the licensee to publish modifications, or that prohibits additional restrictions on downstream use, is incompatible with a proprietary licence that prohibits modification and redistribution. The conflict is resolved at signature or not at all.

Due diligence on the stack

Before a proprietary model is licensed into an environment that also contains an open-weight model, four questions should be answered in writing. Which models are in the stack, at which versions. Which licence governs each. Whether the licence permits the intended commercial use at the intended scale. And whether the licensee’s own obligations to its customers are consistent with the licence terms.

The answers belong in a schedule, with a compliance allocation clause that says who is responsible for each licence and an indemnity for breach. A licensor that discovers an incompatible open-weight component after the product ships has two bad options: withdraw the product or accept the risk.

Open weights also interact with trade secret protection. An open-weight model in the stack does not destroy the secrecy of the licensor’s own weights while they remain separate and controlled. It does destroy the argument if the licensor’s weights are merged with the open model and the result is published, which is why the derivative provisions and the open-weight schedule have to be read together.

Data, outputs and prompts

Two data questions arise in a model licence, and they run in opposite directions.

The first is what the model processes on the Chinese side. If personal information is processed in China, PIPL applies to the Chinese operator, including the cross-border transfer rules if data leaves the country. The licensor may be a joint controller or a processor depending on the arrangement, and that characterisation should be documented rather than assumed.

The second is what training data was used and whether it can lawfully be transferred. Where the licensor’s training data includes personal information, the transfer of weights derived from it raises questions that a simple licence does not answer. The practical answer is a disclosure schedule: the categories of data used, the sources, the lawful basis relied on, and any restrictions that the licensor’s own suppliers imposed. A licensee that discovers the provenance problem after signature will seek an indemnity, and an indemnity for a problem the licensor knew about is expensive.

A data schedule addressing purpose, roles, storage location, retention and cross-border mechanics is now a standard part of a model licence, not an optional annex.

Outputs

The licensee will want to use outputs commercially, including in its own products. The licensor will want to prevent outputs being used to train a competitor. A clause that gives the licensee ownership of outputs but prohibits their use for training a competing model achieves both, provided the prohibition is enforceable and monitored.

Enforcement of that prohibition is the weak point. It cannot be checked by inspecting weights, because the competing model is trained elsewhere and looks different. What can be checked is volume and pattern: a licensee that generates far more outputs than its product requires, or generates outputs in the shape of training data, is a licensee worth auditing. The licence should therefore include a volume anomaly clause, permitting the licensor to require an explanation when usage exceeds a stated multiple of the licensee’s reported product volume.

Prompts

Prompts are a separate category. They may contain the licensee’s confidential information, personal information or trade secrets. The licensor’s right to log and use prompts for model improvement should be stated, and the licensee’s right to opt out should be addressed. Where prompts contain personal information, the logging creates a data compliance obligation that must be allocated.

The commercial default that Chinese enterprise licensees now expect is no logging of prompt content, with aggregate usage metadata only. A licensor that intends to log content for evaluation should say so at term-sheet stage, because the point is usually a procurement blocker rather than a negotiable term.

The Chinese regulatory overlay

A model used internally in a factory is not the same as a model offered to the Chinese public.

Where the licensee offers a generative AI service to the public, the service may require filing with the Cyberspace Administration of China and, where the model has public opinion or social mobilisation attributes, a security assessment. Algorithm filing may also be required for the recommendation or generation functionality. The agreement should identify which party is the filing entity, who prepares the materials, who bears the cost and what happens if filing is refused. A launch that depends on a filing the licensee did not know it needed is a launch that slips by months.

Where the licensor provides the model from outside China and the licensee operates the service, the regulatory position is that the licensee is the service provider. The licensor’s obligations are contractual and, where it processes data, may also be regulatory.

The sequence below is the one that determines whether a launch date is achievable.

StepAuthorityTypical timingWhat foreign companies miss
Classify the service: internal tool or public-facing serviceCounsel, before launchAt design stageThe classification follows the users, not the technology
Filing for a generative AI serviceCyberspace Administration of China, at national and local levelBefore public launchMaterials describe the model and its training data, which the licensor may not wish to disclose
Security assessment where the model has public opinion or social mobilisation attributesCyberspace Administration of ChinaBefore public launchThe assessment is a substantive review, not a formality
Algorithm filing for recommendation or generation functionalityCyberspace Administration of ChinaBefore or alongside launchThe filing entity is the operator, which is usually the licensee
Data transfer route for any data leaving ChinaCyberspace Administration of China, through security assessment, standard contract or certificationBefore the first transferThe route is chosen by volume and sensitivity, and the filing takes longer than the product build

The timings are indicative and the practice varies between provinces. What does not vary is that the filing entity is identified by the facts of who operates the service, not by what the contract says, so the contract should allocate the work to the party the regulator will look to.

Termination, deletion and proof

Deletion is the hardest obligation to verify and the easiest to breach.

A credible deletion clause requires the licensee to delete all copies of the weights, including backups, development environments and derived checkpoints, and to certify the deletion. The licensor should have the right to verify by technical means, such as a licence-server check, a watermark test or an inspection of deployment logs, and the certification should be backed by a liquidated damages provision.

The clause should also cover the artefacts around the weights: the inference code, the configuration, the tokeniser, the evaluation harness and the documentation. A licensee that deletes the checkpoint and keeps the evaluation suite has kept a substantial part of the package.

Where the model has been embedded in shipped products, deletion is impossible, and the agreement should instead provide for a sunset: the licensee may continue to support existing devices for a defined period, may not deploy the model in new products, and must pay a continuing royalty for the installed base.

Two obligations complete the exit and are usually omitted. A transition obligation, under which the licensor returns or destroys the licensee’s data, provides reasonable assistance to migrate to a replacement model for a stated period, and delivers the licensee’s own derivative artefacts where those belong to the licensee rather than to the licensor. And a survival clause that keeps confidentiality and audit rights alive for a defined period after termination, long enough to audit the final royalty period and to pursue a leak that becomes visible only later. An exit that ends at the deletion certificate leaves the licensor unable to prove anything about the last two years.

Pricing and metering

Model licences are priced in four common ways, and the choice affects the control regime.

A one-off licence fee suits an on-premise deployment with a defined scope, because the licensor is paid for the grant and the control burden is fixed. A subscription suits a hosted service, because it aligns payment with continuing access and gives the licensor a switch. A per-inference or per-device royalty suits embedded deployment, because it tracks usage and scales with the licensee’s success. A revenue share suits a partnership where the licensee’s commercial outcome is uncertain.

Each model needs a different audit mechanism. A subscription needs access logs. A per-inference royalty needs metering that the licensor can verify. A per-device royalty needs serial-number reporting and, where the device is exported, customs data. A revenue share needs financial statements and a right to audit them.

The pricing model should be chosen with the control mechanism in mind, because a royalty that cannot be measured is a royalty that will be underpaid. A licensor that accepts a per-inference royalty on an air-gapped deployment has accepted a royalty based on the licensee’s own count, and should price that risk into the rate.

Practical checklist

The economics of a model licence are decided in the schedule and the delivery clause, not in the royalty rate. A licensor that gets the first two right can price the third with confidence; one that gets them wrong is selling an asset it will not be able to identify in a dispute.

Geason Law Offices advises foreign model owners and Chinese deployers on the structure of model licences, the technical controls that support them and the filings that precede a launch. The technology licensing service page describes how we approach the transaction, and the related guides on the scope clause in a China technology licence and contributing IP to a Chinese joint venture cover the two adjacent questions that most often arise when a model is the asset being transferred.

Frequently Asked Questions↑ Back to article

What exactly is being licensed in an AI model licence?

Usually several distinct assets with different legal characters: the model weights or parameters, the inference code, the training pipeline and data, the documentation, and the trade marks. Each has a different default rule and a different enforcement path, so the agreement should schedule them separately rather than treating them as one item called the Model.

Are model weights protected as trade secrets in China?

They can be, where they are not public, have commercial value and are subject to corresponding confidentiality measures. That means access controls, marking, individual undertakings and a disclosure log. Weights that are downloadable from a public repository are not secret, and a licence that treats them as such will not survive scrutiny.

Is on-premise deployment riskier than an API licence?

Yes, materially. An API keeps the weights on the licensor's infrastructure and limits what the licensee can extract; an on-premise deployment places the weights in the licensee's environment, where copying is technically trivial and detection is difficult. On-premise licences need stronger audit, watermarking and contractual controls, and often a higher price.

Can we prohibit fine-tuning and distillation?

Yes, contractually, and it is worth doing expressly. A general restriction on reverse engineering may not obviously cover distillation or the training of a competing model on the licensed model's outputs. Name both, and add a restriction on using outputs to train a competing model, which is the commercially significant leakage route.

What data obligations apply when a Chinese partner runs the model?

If personal information is processed in China, PIPL applies to the Chinese operator, including the cross-border transfer rules if data leaves the country. If the model processes data from the licensor's home market, the licensor's own law applies. The two regimes should be reconciled in a data schedule before launch rather than after a regulator asks.

How do we verify the licensee has deleted the weights on termination?

With a combination of contractual and technical measures: a written certification of deletion, a right to audit the licensee's systems, and technical evidence such as deployment logs, licence-server records or watermark checks. Certification alone is weak; a right to inspect and a technical trail make it credible.

Does exporting model weights from China require approval?

It depends on the model and on the applicable Chinese rules, which continue to develop. Where the weights are treated as controlled technology, export control may apply; where the model is a generative AI service offered to the public, filing and registration obligations arise separately. The classification should be settled before the transfer, not after.

Geason Law Offices AI Industry Research Group | Reviewed by He Qianggao (Lawyer License No. 14403201310092337, 10+ years of practice)