Your weights sit on a cluster in Shenzhen, or in a cloud account registered in Shanghai, and a senior machine learning engineer has just resigned to join a competitor. Six months later a product appears that behaves uncomfortably like yours. You ask your lawyer whether the model is patented. It is not, and that is not necessarily bad news. In China, as elsewhere, the asset that is hardest to copy is the one nobody can see, provided you can prove it was secret and that you treated it that way.

That proviso does most of the work in this area. Trade secret protection in China is not a registration system that grants a right and then defends it. It is a fact pattern that a court either accepts or rejects, assembled from records that existed before the dispute. The law asks whether you took corresponding measures to keep the information confidential, and the answer is found in access logs, key management, contracts, training records and the small decisions made when someone asked for an exception.

Most companies that lose a weights case do not lose it on the definition of a trade secret. They lose it because the evidence of their own security programme is thin, inconsistent, or dated after the leak. The measures described below are the ones that change that outcome, and the section on what destroys secrecy is the one to read first if the model is already in production.

Who this guide is for

Stage: Execute. Use this when the model already exists and the question is how to keep it, not whether to build it.

The foreign element adds three complications that a purely domestic Chinese company does not face. The asset is often held by an entity that is not the entity that employs the people who can reach it. The parent needs access for oversight, which creates a cross-border data flow. And the contracts that matter are frequently governed by foreign law while the conduct happens in China, so the confidentiality obligation and the enforcement mechanism sit in different legal systems. Each is manageable. None is managed by a standard template.

Why model weights fit trade secret better than patent

Chinese patent law rewards disclosure. A specification is published, typically eighteen months after filing, and the monopoly lasts twenty years from filing. Trade secret law rewards silence: protection lasts as long as the information stays secret, and there is no registration, no examination and no publication.

DimensionInvention patentTrade secret
DisclosureFull specification publishedNone
DurationTwenty years from filingIndefinite while secrecy holds
Cost and timingSubstantive examination, typically two to three yearsImmediate; cost is operational
Scope of rightBlocks independent creationReaches only improper acquisition, use or disclosure
DetectabilityClaims are readable, infringement can be assessedInfringement is hard to detect
Reverse engineeringCovered if claims read on the productLawful reverse engineering is a defence
Best fitTraining methods, inference architecture, hardware schedulingWeights, adapters, datasets, hyperparameters, pipelines

Weights are a poor patent subject. They are high-dimensional, opaque and difficult to describe as a technical feature, and a patent on a training method does not stop anyone from using a checkpoint you trained. Publishing a specification destroys secrecy permanently, so the decision is one-way.

The one-way character is worth pausing on. A company can keep a method secret and later patent a different, improved version. It cannot patent a method, publish the specification and then recover trade secret protection for the same information. The disclosure is irrevocable, and it is worldwide. Companies that publish for reputational reasons — a paper at a conference, a technical blog post, an open benchmark submission — are making an intellectual property decision with marketing consequences, and the decision is usually taken by someone who does not see it that way.

The workable architecture is a split. Patent the method where it produces a technical effect, such as a training technique that improves accelerator utilisation or a compression method for edge deployment, and keep weights, adapters, evaluation sets and data pipelines as trade secrets. Draw the boundary deliberately: papers and patent specifications describe architecture, never checkpoints or the data that produced them.

Drawing the boundary means deciding, in advance, what the company will say publicly about how the model was built. A paper that describes the architecture, the loss function and the evaluation protocol but not the data composition, the hyperparameter schedule or the checkpoint selection criteria preserves most of the commercial value. A paper that publishes the full recipe in the name of reproducibility has given away the part that competitors find hardest to reconstruct. Both are legitimate choices. Only one of them is usually made deliberately.

Two limits matter. A trade secret does not stop independent development or lawful reverse engineering, and if your weights can be reconstructed from API outputs, the non-public element weakens over time. That makes rate limiting, output perturbation and query monitoring legal controls, not merely engineering preferences.

The three elements, and what counts as a reasonable measure

Under the Anti-Unfair Competition Law (AUCL), a trade secret is technical or business information that is not known to the public, has commercial value, and for which the rights holder has taken corresponding measures to keep it confidential. All three elements must hold.

Non-publicness means the information is not generally known or readily obtainable by relevant people in the field. Internal use, employee access and disclosure to a vendor under confidentiality do not make weights public, provided access is controlled and the circle of recipients is limited. Commercial value is usually straightforward for a model in production, and potential value counts.

Reasonable measures are where cases are won and lost. Chinese courts assess proportionality, not perfection, and they look for a system that was operating before the leak. A signed policy with no enforcement is weak evidence; an access matrix with logs is strong.

LayerTypical measureWhat it evidences
Identity and accessSingle sign-on, multi-factor authentication, least privilege, per-project rolesAccess was controlled and attributable to individuals
Model protectionEncryption at rest and in transit, customer-managed keys, no local copies of weightsTechnical barriers existed
TraceabilityCanary outputs, fingerprinting, query and export logging, build hashesLeakage can be detected and traced
ContractConfidentiality clauses, IP assignment, vendor terms, no-training commitmentsRecipients were under an obligation of secrecy
Physical and networkSegregated training environments, device controls, no removable mediaCasual exfiltration was restricted
GovernanceAsset register, classification policy, periodic audits, exit checklistThe regime was systematic rather than ad hoc

What a measure has to prove

The table above describes categories. A court, and a counterparty’s lawyer before any court is involved, applies a narrower test to each measure. Five attributes decide whether a measure counts.

It existed before the event. A policy adopted after the engineer resigned is evidence of reaction, not of a regime. Documents should carry dates, and the dates should precede the disclosure they are meant to protect. Where a measure was introduced late, say so and explain what preceded it, because an honest chronology is more persuasive than a backdated policy that the other side can date from metadata.

It is specific to this asset. A general confidentiality clause covering “all company information” is a starting point, not proof that weights were treated as secret. The classification scheme should name model artefacts, adapters, evaluation sets and training configurations as categories, and the access rules should apply to them by name.

It was applied to this person. The strongest evidence in a departure case is a record showing that the individual was granted access to a named project, acknowledged the terms, and was subject to a control that would have detected bulk extraction. A signed company-wide policy that nobody can show was communicated to the individual is weaker than a single access approval email with a named approver.

It generated a record. Measures that leave no trace are difficult to prove. An access control that logs nothing proves intent but not operation. Prefer controls that produce evidence as a by-product: authentication logs, repository audit trails, key-use records, export tickets, device check-in records.

It was enforced. One refused exception is worth a page of policy. If an engineer asked for a local copy of the weights and was refused, or asked for a personal cloud account and was directed to the sanctioned route, the refusal is evidence that the regime was real. Companies should keep these refusals rather than treat them as routine administration.

Proportionality runs through all five. A five-person start-up is not expected to run a hardened facility. It is expected to have done something that a reasonable observer would recognise as an attempt to keep the weights secret, applied consistently, and documented at the time.

The measures, layer by layer

The categories in the table above fail in predictable ways. The account below sets out what works at each layer and what usually goes wrong, because the failure modes are more instructive than the labels.

Identity and access. Single sign-on with multi-factor authentication, per-project roles and least privilege are the foundation, and they are also where AI teams most often make an exception. A shared service account with an admin key, used by three engineers because provisioning is slow, destroys attribution. When weights leave, nobody can say who had access. The remedy is unglamorous: named accounts, short-lived credentials, a request process that takes minutes rather than days, and an access review each quarter that actually removes people.

Repository and artefact control. Weights should live in a model registry with per-project permissions, not in a shared drive, a git repository with a permissive remote, or an object storage bucket whose access policy was set during a hackathon. The common failures are a bucket with public read enabled for a demo and never closed, checkpoints committed to a repository that is later made public, notebooks with embedded credentials, and export paths that any engineer can use without a ticket. Each of these is discoverable after the fact and each is fatal to the argument that access was controlled.

Encryption and key custody. Encryption at rest with provider-managed keys protects against a stolen disk and not against the provider, an administrator, or a subpoena served on the cloud account. Customer-managed keys held in a key management service, with a separate custodian and an audit trail of key use, change the evidential picture: decryption is an event that someone authorised. Key rotation and separation of duties between the team that holds the data and the team that holds the keys are worth more than a stronger cipher.

Environment segregation. Training, evaluation and production in a single tenant, with contractors in the same project, means every control is shared. Separate environments with separate credentials, and separate projects for external collaborators, limit the blast radius of a single compromised account and make the access record legible.

Logging and data loss prevention. Query logs, export logs, egress monitoring, endpoint controls on removable media, and a rule that weights do not exist on laptops are the controls that produce evidence after a leak. The evidential value lies in the baseline: if the logs show normal behaviour for eleven months and a 40 GB export in the final week, the anomaly is the case. Without a baseline, the same log shows nothing.

Physical and network. A segregated training environment, controlled physical access to the cluster room, visitor records and device lockers matter in proportion to the value of the asset and the number of people with access. For most companies the physical layer is secondary. For a company whose entire model is trained on a single on-premises cluster, it is not.

Contract. Confidentiality clauses, present assignment of work product, vendor terms with a no-training commitment, and intercompany agreements with affiliates. Contracts do not prevent disclosure; they establish that every recipient was under an obligation, which is a necessary element of the claim against a recipient and a necessary foundation for the claim against a new employer.

Governance. An asset register that lists the weights, adapters, datasets, evaluation sets and configurations, with a classification for each; a policy that says who may access which class; a periodic audit; and an exit checklist. Governance is what turns a set of tools into a regime, and it is what a court means by a system.

What destroys secrecy

Destruction comes in three forms, and companies routinely confuse them. Some acts destroy the secrecy of the information. Some destroy the evidence that it was secret. Some destroy the remedy while leaving the secret intact. The response differs for each.

Acts that destroy secrecy. Publishing the training recipe, the data composition or the hyperparameter schedule in a paper or a blog post removes those elements from protection permanently. An unrestricted public demo allows extraction, and extraction repeated at scale erodes the non-public element of the weights themselves. Open-sourcing the training code with configuration files and data manifests gives away the pipeline. A model card that names the dataset, the evaluation set and the training procedure is a disclosure decision dressed as documentation. A conference talk with screenshots of the internal dashboard can disclose more than the speaker intends. Each of these is irreversible, and each is usually taken by someone with no view of the intellectual property consequences.

Acts that destroy the evidence. A shared admin account, an unlogged export path, a bucket without access logs, a laptop that was wiped before anyone imaged it, a departure handled by an HR administrator with no technical involvement. None of these makes the weights public. All of them make the case unprovable, which in practice is the same outcome.

Acts that destroy the remedy. Failing to pay non-compete compensation, so the clause does not bind. Using a contractor without a written assignment, so the company does not own the work. Relying on a base model licence that requires the derivative to be disclosed, so the company cannot claim secrecy in the fine-tune. Continuing to distribute after a licence breach, so a contractual right is lost. These leave the secret intact and remove the instrument that would have protected it.

A single incident often combines all three. An engineer copies the weights to a personal drive, the company discovers it three weeks later, the access logs have rolled over, the non-compete was never funded, and the contractor who built the training pipeline was never assigned the copyright. The weights are still secret in the formal sense. Nothing can be done about them.

Extraction and the decay of secrecy

A model served through an API leaks slowly. Each query returns information about the function the weights compute, and enough queries allow a competitor to train a substitute that behaves similarly on the inputs that matter commercially. The technique is well understood and does not require the attacker to obtain the weights.

Chinese trade secret law reaches improper acquisition, use and disclosure of the secret information. Where a competitor reconstructs behaviour from outputs rather than obtaining the weights, the analysis is harder and the outcome less certain, which is precisely why the engineering controls matter. Rate limiting raises the cost of extraction. Output perturbation, where the served model returns slightly noisy or rounded results, degrades the fidelity of a substitute. Query monitoring detects the pattern — a sudden volume from one account, systematic coverage of an input space, automated timing — and produces the record that supports a complaint.

Fingerprinting serves the same dual purpose. Canary inputs with known expected outputs, distinctive error patterns on edge cases, and versioned build hashes allow you to show later that a competitor’s model derives from yours. The same artefacts that help an engineer debug are the exhibits in a misappropriation case. Companies should design them deliberately rather than hope that behaviour will be distinctive enough.

Distillation by a customer is the quiet version of the same problem. A customer with API access and a commercial incentive can generate a training set from your outputs and train a smaller model. The contract should prohibit it, the terms should define it precisely, and the monitoring should look for the query pattern that precedes it. A prohibition that nobody monitors is a clause, not a control.

People: contracts, non-competes and departures

Employees are the main leakage route. Chinese labour law allows confidentiality clauses, and allows non-competes for senior managers, senior technical staff and others who actually carry confidentiality duties. A non-compete is capped at two years, requires monthly compensation after termination, and must define a reasonable restricted scope, territory and set of competing activities. If compensation is not paid, the clause may not bind, and courts narrow clauses that are drawn to freeze a person out of an entire industry.

Recent judicial guidance has further tightened the use of non-competes against employees who do not genuinely access secrets, so a blanket clause applied to every engineer is a weak instrument. A documented access control regime is worth more than an aggressive clause.

The compensation obligation is a budget item, not a drafting point. Companies that paper every senior engineer with a non-compete and then decline to pay discover, at the moment of enforcement, that the clause is unenforceable and that the money saved was never the constraint. Decide which roles genuinely carry secrets, fund those, and let the rest rely on confidentiality and assignment.

Work product clauses need to be specific. Inventions made by an employee in the course of work generally belong to the employer as service inventions, but the boundary between work and personal projects is contested in AI research, where engineers experiment at home. Define work product to include weights, adapters, code, datasets, prompts and evaluation sets, require a present assignment, impose return and deletion duties on exit, and require notice of conflicting obligations.

“Model” should be defined to include checkpoints, low-rank adapters, embeddings, tokenisers, training configurations and evaluation sets. A definition that covers “software and source code” leaves the weights outside, which is the one asset that matters.

Contractors and data annotators deserve the same treatment, because copyright and ownership in commissioned work may not transfer automatically and annotators see sensitive material. Use written assignment, confidentiality, no-retention and device rules.

The departure playbook matters more than the contract. On the day notice is given, revoke access, preserve logs before they roll over, collect devices, conduct an exit interview and remind the person of continuing obligations. If logs show bulk downloads or exfiltration, move quickly to evidence preservation and consider a police report; delay is the most common reason such cases fail.

The first twenty-four hours deserve a written sequence, because they are chaotic and the decisions are irreversible.

HourActionOwnerWhy it cannot wait
0–2Revoke access to the registry, cluster, cloud accounts and repositoriesIT with the managerEvery further minute is an opportunity to copy
0–4Export and preserve authentication, repository and export logsSecurityRetention periods are short and rotation destroys the record
2–6Collect devices, tokens and any physical media, with a signed receiptIT and HRVoluntary return later is unreliable and looks optional
4–8Interview the manager and the team about recent behaviour and accessHR with legalRecollection fades and colleagues compare notes
6–24Serve the written reminder of confidentiality, return and non-compete obligationsLegalEstablishes the date of notice and the scope of duty
24Decide whether to preserve evidence formally or notify the policeLegal with managementDelay weakens both routes

Two details are commonly missed. Personal cloud accounts and messaging groups are where copies actually live, so the exit checklist should require a written confirmation that work material has been deleted from personal accounts and devices, with the consequence of a false statement explained. And the departure of a manager matters more than the departure of an engineer, because the manager knows which assets exist and where they are held.

Partners, clouds and open-weight fine-tunes

Joint development is where foreign companies most often lose control. Default rules on commissioned and joint work are unfavourable, so the contract should allocate ownership of weights, improvements, evaluation data and the right to enforce. Define “model” to include checkpoints, low-rank adapters, embeddings, tokenisers and training configurations, or you may find that the licence covers only source code.

A joint development clause that allocates “foreground intellectual property” without listing the artefacts produces an argument about whether weights are foreground intellectual property. List them. Allocate the right to enforce separately from ownership, because a co-owner who will not join a claim can block it. And provide for what happens to the weights on termination: who keeps a copy, who must delete, and what happens to a model that neither party can retrain alone.

Cloud providers hold physical access to the asset. Contract for no access to model artefacts, no training on your data, encryption with keys you control, a committed region, audit rights, deletion certification and notice of government access requests. Separately, deployments of certain scale or sensitivity engage China’s Multi-Level Protection Scheme (MLPS) and may require a security assessment by the Cyberspace Administration of China (CAC).

The contract terms that matter are narrower than the standard data processing schedule. Ask four questions. Can the provider’s staff read the model artefacts, and under what approval? Is the customer data excluded from any training or improvement of the provider’s own models? Where are the keys held, and who can use them? What notice will the provider give of a government request, and what will it disclose if it cannot give notice? A provider that cannot answer the third question in writing is a provider whose encryption does not protect you from the provider.

Open-weight fine-tuning creates a quieter risk. Base model licences vary widely: some restrict commercial use, some require attribution, some impose acceptable-use terms that flow to derivatives, and some require you to identify the base model. A fine-tune is often a derivative, so the licence can limit your ability to keep the result secret or to license it on your own terms. Record the base model version and the licence text you relied on, and check whether the licence requires any public statement that would conflict with your secrecy posture.

The conflict is easy to create and hard to unwind. A licence that requires the derivative to display “built with X” makes the lineage public, which may be commercially acceptable and is fatal to a claim that the lineage is a secret. A licence with an acceptable-use condition that flows to derivatives means the fine-tune inherits restrictions you did not negotiate. A licence that permits commercial use below a threshold may be breached by success. Read the licence before the fine-tune, record the version, and keep the text.

Enforcement: injunctions, damages and criminal referral

Civil enforcement has become more effective. Courts can grant behaviour preservation, which functions like an injunction, before trial in urgent cases, and can order a party to stop using the model. Where the rights holder presents reasonable evidence of misappropriation and the defendant controls the proof, the burden can shift to the defendant to explain its lawful source, which is powerful when a new employer claims independent development.

Damages follow a sequence: the owner’s actual loss, then the infringer’s profit, then a reasonable multiple of a licence fee, then statutory damages, which are capped at RMB 5 million. Intentional and serious infringement can attract punitive damages of up to five times the base figure, and reasonable investigation costs are commonly awarded.

Administrative complaints to the market regulation authorities can produce fines and cease-and-desist orders. Criminal referral under the trade secret provisions of the Criminal Law is possible where loss or unlawful gain passes the thresholds set by judicial interpretation, and it is often the strongest pressure point when weights leave with a departing team. The thresholds are monetary and periodically adjusted, so confirm the current figures with counsel.

The hardest element is always proof that the defendant’s weights derive from yours. Build that case technically: canary behaviour on probe inputs, distinctive error patterns, watermarking, and forensic review of repositories where access is available. Secrecy engineering and litigation strategy are the same project.

That sentence deserves emphasis because it is where budgets are wasted. Companies spend on security and separately on litigation, and the two teams never meet. The security team’s logs are the litigation team’s exhibits. The litigation team’s technical expert needs the canary set that the security team designed. A company that builds the evidential trail into ordinary operations spends far less at the point of dispute than one that starts assembling it after the letter arrives.

Cross-border access to China-held weights

If an overseas parent needs to reach weights held in China, treat it as a compliance design problem. Weights are usually not personal information, but the training data often is, and a model that memorises personal information can transmit it through its parameters. Cross-border transfer of personal information requires one of the statutory mechanisms, and important data may trigger a CAC security assessment. Add export control screening where the technology may be restricted, and paper the access route with an intercompany agreement, least-privilege remote access, session logging and a prohibition on local downloads.

The practical design has four parts. Access is brokered through a jump host in China with session recording, so the parent sees the model without holding a copy. The intercompany agreement allocates ownership, confidentiality and liability, and states which entity may instruct whom. Where personal information is involved, the transfer mechanism is documented before the first session. And the export control position is assessed at the level of the technology and the destination, because a model trained on controlled technology can carry the restriction with it.

Documents, filings and timeline

Trade secrets are not registered, so the timetable is set by employment events, deployment decisions and disputes. The steps below are the ones with a fixed sequence.

StepForum or authorityTypical timingWhat foreign companies miss
Asset classification and registerInternalBefore the first production modelWeights left outside the definition of confidential information
Employment and contractor termsInternal, governed by labour lawAt hire and on renewalAssignment covers code but not weights or adapters
Non-compete design and fundingInternal, enforced through labour arbitrationAt hire, funded after terminationCompensation not budgeted, so the clause fails
MLPS grading and filingPublic security authoritiesBefore or at launch, by gradingTreated as an IT project rather than a filing obligation
CAC security assessmentCyberspace Administration of ChinaBefore a triggering transfer or deploymentAssumed to apply only to consumer apps
Evidence preservationNotary, court or policeImmediately on suspicionLogs allowed to rotate before anyone looks
Behaviour preservation applicationPeople’s courtBefore trial, in urgent casesFiled after the damage is done
Criminal referralPublic security authoritiesAfter loss is quantifiedThresholds assumed rather than checked

Cost and commercial impact

The cost of a secrecy programme is mostly operating expense, and it is small relative to the asset. Identity infrastructure, key management, logging and an access review process are things a company of any size already buys in some form; the incremental cost is configuration and discipline rather than procurement.

Three items do carry real budget. Non-compete compensation for the roles that genuinely hold secrets is a recurring monthly cost after termination, and it has to be funded before the clause is relied on. MLPS grading and any CAC assessment involve consultants and engineering time, with the cost driven by the grading level and the complexity of the system. Evidence preservation and forensic work are cheap when done routinely and expensive when done in a hurry.

The litigation side is where the numbers become large and uncertain. Damages are computed from loss, profit, a licence benchmark or the statutory cap, and the cap of RMB 5 million with punitive multiples up to five times is meaningful but not transformative. The larger commercial effect is on the business: an injunction that stops a product, a customer who learns of the dispute, and an investor who re-prices the asset. Companies settle weights disputes for reasons that have little to do with the damages calculation.

The asymmetry worth understanding is that the defendant usually has the evidence. The departing engineer’s new employer knows what it built and how. That is why the burden shift matters, and why the technical fingerprints matter: they are the only way to make the defendant’s own records work against it.

What to do next

  1. Classify assets now: weights, adapters, datasets, evaluation sets, configurations and prompts, with a clear decision on which are secret and which are patentable.
  2. Write a short secrecy policy and then evidence it through an access matrix, multi-factor authentication, encryption and logging.
  3. Audit employment and contractor contracts for work product assignment, confidentiality, return obligations and enforceable non-competes with a funded compensation budget.
  4. Map every external party with access, including cloud vendors, annotators, integrators and overseas affiliates, and paper each one with confidentiality and no-training terms.
  5. Fingerprint the model with canary outputs, versioned build hashes and probe sets you can use for later comparison.
  6. Review the base model licence before every fine-tune, and record the version and terms you relied on.
  7. Run a departure drill: define exactly what happens in the first twenty-four hours after an ML engineer resigns.
  8. Prepare the litigation file before you need it, covering secrecy measures, access logs, a loss model and a technical expert.
  9. Set retention periods for access and export logs that exceed the likely discovery window, and test that they hold in practice.
  10. Decide what the company will publish about the model, and route that decision through whoever owns the intellectual property position.

Where this leads

The employment side of the problem is covered in more detail in AI Company Employment Boundaries and Cross-Border Employment and Tax for AI Companies, which address the contracts that have to hold before a departure happens. Where the model was built on someone else’s weights, Open-Source Licence Compliance explains how the base licence constrains what can be kept secret. Companies that have not yet settled where the team and the model sit should begin with Do I Need a Chinese Entity?.

For advice on a specific model, contract or dispute, see the intellectual property protection service page or contact the firm.

Frequently Asked Questions↑ Back to article

Can model weights be protected as trade secrets in China?

Yes, provided the weights are not publicly available, have commercial value because they are secret, and you have taken reasonable confidentiality measures. The third element is where most claims fail, because it requires documented controls such as access restrictions, encryption, logging and contractual obligations rather than a policy that exists only on paper.

Why not patent an AI model instead of relying on trade secrets?

A patent requires full disclosure and expires, and it is difficult to claim high-dimensional weights in patent terms. Patents are better suited to training methods, inference architectures and hardware scheduling, where a technical solution can be described and detected. Most companies patent the method and keep the weights and data secret, with a deliberate boundary between what is published and what is not.

What confidentiality measures do Chinese courts actually expect?

Courts look for a proportionate, functioning system rather than perfect security. Typical expectations include role-based access control, multi-factor authentication, encryption in transit and at rest, query and export logging, confidentiality clauses with everyone who has access, and records showing the controls operated before the leak. Measures created after the dispute carry little weight.

Are non-compete agreements enforceable against AI engineers in China?

Only within limits. Non-competes are available for senior managers, senior technical staff and others with genuine confidentiality duties, are capped at two years, and require monthly compensation after termination. Courts strike down clauses that cover junior staff without access to secrets, and recent judicial guidance has tightened that scrutiny, so access control matters more than a broad clause.

Who owns a fine-tuned model built on an open-weight base model?

Ownership depends on the base model licence, your development contracts and whether the fine-tune is treated as a derivative work. Many open-weight licences impose conditions such as attribution or acceptable-use restrictions that carry into derivatives, and some require you to identify the base model. Check the licence before you treat the fine-tuned weights as a proprietary secret.

How are trade secret damages calculated in a Chinese lawsuit?

Courts generally start with the owner's actual loss, then the infringer's profit, then a reasonable multiple of a licence fee, and fall back on statutory damages where none of these can be established. Intentional and serious infringement can attract punitive damages of up to five times the base amount. Legal costs and reasonable investigation expenses are commonly awarded to a successful claimant.

Is it safe for our overseas headquarters to access model weights stored in China?

It can be, but it needs structure. Weights themselves are usually not personal information, yet the training data behind them often is, and cross-border transfer rules under PIPL and the Data Security Law can apply to that data or to memorised content. Use least-privilege remote access with logging, an intercompany agreement, and a transfer mechanism where personal information is involved.

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