China AI Export Controls: Open-Weight Model Rules
China's draft AI export controls make API access and self-hosting different compliance choices. This buyer guide explains the thresholds, the licensing routes, and what to ask suppliers in writing.
The Short Answer
On October 16, 2026, China AI export controls moved from rumour to a published draft, and every business that buys, builds on, or hosts a Chinese open-weight model should read it as a procurement document, not just as a news story. The draft, published by China's Ministry of Commerce and Ministry of Science and Technology…
On October 16, 2026, China AI export controls moved from rumour to a published draft, and every business that buys, builds on, or hosts a Chinese open-weight model should read it as a procurement document, not just as a news story. The draft, published by China's Ministry of Commerce and Ministry of Science and Technology as Announcement No. 48, proposes a revision of the catalogue of technologies prohibited or restricted from export. It covers 40 technologies in total, and 23 of them are in AI, including large models, agents, computing clusters, and chip design. Comments are open until November 15, 2026, and one interim step, pre-export registration, has applied since October 16. This article explains what the draft says, why API access and self-hosting are now different compliance choices, and what to ask a supplier before you commit.
One point matters first. This is a draft, not final law, and thresholds, licensing, and definitions could still change before a final catalogue takes effect. Nothing here is legal advice.
What China AI export controls actually cover
The draft does not ban Chinese AI, and it does not close off open-weight models. It draws a line around the largest and most sensitive technologies and says that some of them need permission before they leave China. For large models, controls attach to what the draft calls the core technology: the weights, the key checkpoints, the underlying source code, the architecture design, and the training and inference methods. That list matters because it tells you what is actually controlled. It is not the chat window, and it is not the finished app. It is the model file and the know-how that built it.
A large model is restricted if it crosses any one of three lines at the time of export: training compute of at least 10^26 floating-point operations, at least 2 trillion parameters, or capability at a global leading level, defined as a top-10 place in a public global ranking by a major Chinese or foreign evaluator or a large industry platform, judged on overall performance, specific skills, global usage, safety, or training efficiency. Meeting just one of the three is enough.
| Threshold | What it measures | Buyer takeaway |
|---|---|---|
| 10^26 FLOPs of training compute | Scale of training | Only the very largest models cross this line |
| 2 trillion parameters | Size of the model | Very large open-weight releases are most likely caught |
| Global top-10 ranking | Public standing at export | Classification can change as rankings change |
Beyond large models, advanced agents, AI chip design at 14 nanometres or below, and technology behind computing clusters of more than 10,000 accelerator cards are among the restricted items. A smaller set faces a full export ban, including tools that can automatically write or alter malware aimed at critical infrastructure, and large models trained on what China calls its core data and then offered as a service. Most ordinary business AI remains outside these controls. There is also an important exclusion: technology already in the public domain through open source or published academic papers is outside the draft's scope, so weights already released publicly are not caught by the new controls.
China AI export controls: API access vs self-hosting
Here is the distinction that changes buying decisions. Under the draft, confirmed in the ministries' own explainer, using a restricted model through an API, an app or website, a SaaS product, or a lightweight on-device version is not treated as an export. Only overseas local deployment of the restricted technology needs a licence. In plain language, calling a model that stays on the vendor's servers is treated differently from taking the weights home and running them on your own servers.
That split creates two very different compliance paths. If you use a model through an API, the export question is largely the supplier's to manage. If you want the weights themselves, whether to self-host, fine-tune on your own infrastructure, or embed in a product you ship, you are asking for the controlled item itself, and the licence question lands directly on your project. For businesses that chose open-weight models for control, independence, and data residency, self-hosting is now a compliance decision as well as an engineering one.
| How you use the model | Treated as an export? | What it means |
|---|---|---|
| API calls to a vendor hosted model | No, for restricted models | Lowest export friction, vendor dependent |
| SaaS app or website built on the model | No, for restricted models | Vendor controls the deployment |
| Lightweight on-device version from the vendor | No, for restricted models | Lightweight is not precisely defined |
| Downloading weights and running them yourself | Yes, if restricted | Licence needed, build it into your timeline |
| New open-weight release above a threshold | Yes, fast-track route | Automatic licensing, process not yet detailed |
The draft gives open-source releases a fast-track automatic licensing route, signalling it is not designed to end open-weight releases, but to put a registration and licensing step in front of them. How automatic that process proves to be in practice remains an open question.
Why the licence attaches to the model file
The original insight in this draft is easy to miss because it sounds technical, but it rewrites procurement. A licence attaches to the model file, not to the brand, not to the demo, and not to the feature list. Two products can look identical in a sales presentation and sit on opposite sides of the control line, because one ships you weights and the other only sells you access. A procurement checklist that only asks about features, accuracy, latency, and price is now incomplete. It must also ask: what exactly is being transferred, where will it run, and who holds the licence?
Most AI buying checklists were written for a world where the main risks were performance, security, and vendor lock-in. Those risks are still real, but transfer risk has joined them. If your plan depends on receiving weights and hosting them yourself, a licensing delay or refusal is a project risk, just like a missed delivery date. If your plan only depends on API access, that risk largely stays with the vendor. Naming which plan you are on, in writing, at the start of a purchase is now one of the most valuable things a buyer can do.
What to ask a supplier in writing
Do not accept verbal reassurance. Ask for answers in writing, keep them with the contract, and ask again if the model version changes.
- Has this exact model version been registered under the interim pre-export registration in force since October 16, 2026, and can you share the reference?
- How is the model classified: restricted, prohibited, or outside scope, and which threshold, if any, does it cross?
- What is being transferred: API access only, a lightweight on-device version, or the full weights, checkpoints, and source code?
- If weights are transferred, has a licence been granted, applied for, or covered by the automatic open-source route, and what is the timeline?
- Where will the weights sit: in your data centre, in our cloud account, or on our own servers, and who controls each copy?
- If the model is updated mid-contract, who re-checks the classification, and who bears the cost of any delay?
- What happens if a licence is delayed, refused, or withdrawn: do we keep running, do we pause, and who decides?
What to write into the contract
Questions are only useful if the answers become obligations. These clauses belong in any contract involving a restricted model, and they are not a substitute for legal advice.
- Classification warranty: the supplier warrants the model's classification at delivery, and must notify you promptly if a new version, ranking change, or rule change alters it.
- Licence responsibility: the contract states who applies for any licence, who pays for it, and what happens if it is delayed or refused.
- Licensing buffer in dates: milestones should allow time for registration and licensing, so a delay is a managed event, not an automatic breach.
- Termination rights: if a licence is refused, or a final rule prohibits the transfer, you should have a clear right to pause, switch to API access, or exit.
- Continuity for API users: the contract should cover service continuity, notice periods, and data export, so a change in the supplier's position does not strand your product.
- Change control: no new version that crosses a threshold should reach your environment without a fresh classification check and your written agreement.
- Audit and evidence: the supplier should keep, and share on request, the registration and licensing documents behind its claims.
Where the weights sit, and why it matters
Location is the quiet variable here. A model in the vendor's data centre, accessed through an API, keeps the controlled technology in one place. A model on your own servers or on devices you ship moves it somewhere new, under new control. The draft treats those situations differently, and your architecture should reflect that.
This is also where data residency and export control meet. Many businesses choose self-hosting because they want data to stay in country, or full control over uptime and tuning. Those are good reasons, and this draft does not remove them. It simply adds a step: if self-hosting means receiving restricted weights from China, the licence step must be planned, documented, and written into the timeline. Businesses weighing self-hosting should read our guide to sovereign AI at home and self-hosted models for business alongside this article, because the operational case and the compliance case now need to be decided together.
What this means for buyers in the UK, the USA, and Pakistan
The draft is Chinese law in the making, but its effects will reach buyers everywhere, because so many open-weight models originate in China or build on Chinese research. UK and US buyers should note that this draft sits alongside their own export and AI governance regimes, and it does not replace them. Pakistan based buyers, including software houses and startups building for overseas clients, should pay particular attention to self-hosting, because a product built on weights hosted in Pakistan is exactly the kind of overseas local deployment the draft is concerned with.
For all three audiences, the advice is the same. Start with API access where it meets your needs, as the lowest friction route. Move to self-hosting deliberately, with classification and licensing in writing, where control or residency genuinely requires it. Comments close on November 15, 2026, and the final catalogue could refine thresholds and licensing. Our buyer guide to Kimi K3.1 Pro and open-weight AI shows how to evaluate a large open-weight release, while our guide to in-country AI processing and data residency in 2026 covers where your data is processed. The draft adds the compliance layer on top of both.
Frequently asked questions
Is this draft already law?
No. Announcement No. 48 is a draft revision for public comment, with comments closing on November 15, 2026. The one step already in force is interim pre-export registration, in force since October 16, 2026. November 15 is the end of the comment period, not an enforcement date.
Does using a Chinese AI model through an API count as an export?
For restricted models, no. The ministries' explainer states that offering a restricted model through API calls, an app or website, a SaaS product, or a lightweight on-device version is not within the scope of large-model technology export control. Only overseas local deployment needs a licence.
Can I still self-host a large Chinese open-weight model?
The draft does not ban self-hosting in general, but if the model is restricted, transferring its weights for overseas local deployment needs a licence. New open-weight releases above a threshold get a fast-track automatic licensing route. Get the classification in writing from your supplier.
What if the model I already use is already public?
Technology already in the public domain through open source or published academic papers is outside the draft's scope. Weights already released publicly are not caught, although a future version or a new deployment may still need checking on its own facts.
Which models are actually caught by the thresholds?
Only the largest models: at least 10^26 FLOPs of training compute, at least 2 trillion parameters, or a global top-10 ranking at the time of export. Most everyday business models sit well below all three, which is why checking the specific model matters.
What should I do first if my product depends on a Chinese open-weight model?
Ask your supplier, in writing, for the model's classification, its registration status, and whether your deployment involves a transfer of weights or only API access. Then reflect the answers in your contract. If you need help with that assessment, see our AI and machine learning services for how we support businesses choosing and deploying AI.
If you are planning an AI deployment and want the compliance questions answered before you commit to an architecture, contact Skills Technology Solutions and tell us whether you are leaning towards API access or self-hosting. We will help you weigh the options, ask the right questions of your supplier, and choose a setup that stays workable whatever the final rules bring.
Related Articles
Keep reading — more guides you may find useful.
Related Services
Services connected to this guide.
