If you’re a developer outside China, there’s a good chance you’ve used — or at least evaluated — models like Qwen, DeepSeek, or GLM. They’re competitive, often open-weight, and in some benchmarks they match or beat Western alternatives.

May 2026 update: China’s AI content labelling rules (mandatory dual labelling with visible markers + embedded metadata) have been in effect since September 2025. For the full breakdown of labelling requirements across Asia, see AI Content Labelling Laws in Asia. Also new: a Hangzhou court ruled that companies cannot fire workers to replace them with AI — AI implementation is a voluntary business decision, not grounds for termination.

But China has its own AI regulatory framework. And it’s not a single law — it’s a layered system of overlapping regulations, registries, and content rules that Chinese AI providers must navigate.

The question for you: does any of this affect international developers who just want to use these models?

Short answer: it depends on how you use them. This guide breaks it down.

China’s AI regulatory framework

China doesn’t have one “AI Act.” Instead, it has built a stack of regulations over several years, each targeting a different aspect of AI:

1. Algorithm Recommendation Measures (2022) — governs recommendation algorithms and personalization systems. Requires registration of algorithms that influence public opinion or have “mobilization capacity.”

2. Deep Synthesis Provisions (2023) — targets deepfakes and synthetic media. Providers must label AI-generated content and implement real-name verification for users.

3. Generative AI Administrative Measures (2023) — the most relevant regulation for LLM providers. Covers any service that generates text, images, audio, or video. Requires security assessments before public launch.

4. Amended Cybersecurity Law (effective January 1, 2026) — adds explicit AI compliance provisions, increases penalties for violations, and strengthens the legal basis for enforcement across all the above regulations.

These regulations overlap and reinforce each other. Together, they create what’s sometimes described as a “2-3-4-3-2” registration system: 2 AI service categories, 3 foundational regulations, 4 public registries, 3 review processes, and 2 disclosure channels. It’s bureaucratically dense — but that’s China’s problem, not yours. Mostly.

What Chinese AI providers must do

If you’re Alibaba (Qwen), DeepSeek, or Zhipu AI (GLM), operating AI services in China means:

• Registration before launch. AI services must be filed with the Cyberspace Administration of China (CAC) before they go live. This includes algorithm registration and security assessments.
• Content restrictions. Generated content must support “core socialist values.” Content that endangers national security, promotes separatism, or undermines social stability is prohibited. This is why Chinese chatbots behave differently on politically sensitive topics.
• Data localization. Personal data and “important data” collected in China must be stored on servers within China. Cross-border data transfers require security assessments.
• Labeling and traceability. AI-generated content must be identifiable. Providers must maintain logs and be able to trace outputs.

These obligations fall on the provider operating the service in China. They shape how Chinese companies build and deploy their models domestically.

What this means for developers outside China

Here’s the critical distinction: China’s AI regulations primarily target providers operating services within China. They regulate the domestic deployment of AI, not the global use of open-source model weights.

If you’re a developer in the EU, US, or elsewhere, the regulatory picture depends entirely on your deployment model:

• Self-hosting open-weight models → China’s rules don’t apply to you
• Calling Chinese provider APIs → data flows to China, and that’s where it gets complicated

Let’s look at both paths.

The safe path: open-weight models, self-hosted

Models like Qwen 3.6 and DeepSeek R1 are released under permissive licenses (Apache 2.0 or similar). You download the weights, run them on your own infrastructure, and no data leaves your environment.

In this scenario:

• No data flows to China. Your prompts, user data, and outputs stay on your servers.
• China’s content restrictions don’t apply. You’re not operating a service in China. The model weights are just math — the content policies are enforced at the service layer, which you control.
• China’s registration requirements are irrelevant. You’re not launching an AI service in China.
• Your obligations are determined by YOUR jurisdiction — GDPR, the EU AI Act, state-level US privacy laws, etc.

Self-hosting is the cleanest path. You get the model capability without any regulatory entanglement with China. This is why running models locally has become the default recommendation for teams with privacy requirements.

The risky path: API access to Chinese providers

If you call DeepSeek’s API or use a Chinese provider’s hosted inference, your data travels to servers that may be located in China. This creates real problems:

• Data localization in reverse. China’s laws may require that data received by Chinese servers stays in China. Your user data could become subject to Chinese data retention rules.
• GDPR cross-border transfer issues. There’s no EU adequacy decision for China. Transferring personal data to China requires additional safeguards (SCCs, transfer impact assessments) — and even those may not be sufficient given China’s national security data access laws.
• Opacity. Chinese providers’ data processing practices are often unclear to international users. Privacy policies may not meet GDPR transparency requirements.

For most international developers, API access to Chinese providers is a compliance risk that’s hard to mitigate.

GDPR and Chinese AI models

The GDPR friction is already playing out in practice:

• Italy blocked DeepSeek entirely in January 2025, with the Garante (Italian DPA) citing insufficient transparency about data processing and the lack of a legal basis for processing Italian users’ data.
• Belgium, France, and Ireland opened investigations into DeepSeek’s data practices in early 2025.
• These actions target the hosted service — the API and web interface — not the open-weight model files themselves.

The pattern is clear: European regulators treat Chinese AI services the same way they’d treat any non-compliant data processor. The model weights aren’t the problem. The data pipeline is.

If you’re building for EU users, check our guide to GDPR-compliant AI APIs and the broader AI privacy laws overview.

Practical recommendations

If you’re using Chinese open-weight models (Qwen, DeepSeek, GLM):

1. Self-host. Download the weights, run them on your infrastructure (or a trusted cloud provider in your jurisdiction). This sidesteps both Chinese regulation and cross-border data transfer issues.

2. Don’t use Chinese-hosted APIs for production workloads involving personal data or EU users. The GDPR risk is real and enforcement is active.

3. Document your deployment model. If regulators or clients ask, you need to show that no data flows to China. “We self-host the model on EU infrastructure” is a clear answer.

4. Apply your own content policies. Open-weight models don’t come with China’s content restrictions baked in, but they also don’t come with any safety layer. You’re responsible for output filtering in your jurisdiction.

5. Watch the EU AI Act. The EU AI Act’s obligations apply based on where you deploy and who you serve — regardless of where the model originated. Open-weight models from China are treated the same as any other model under EU law.

6. Track regulatory changes. China’s amended Cybersecurity Law (2026) signals continued tightening. If you’re evaluating Chinese API services, reassess regularly.

The bottom line

China’s AI regulations are complex, but for most international developers they’re someone else’s problem — as long as you self-host. The moment data flows to Chinese servers, you inherit a tangle of cross-border compliance issues that are difficult to resolve.

The open-weight releases from Qwen, DeepSeek, and GLM are genuinely useful. Use them. Just run them yourself.

Related links

• Qwen 3.6 Complete Guide
• Qwen 3.6 35B-A3B Complete Guide
• AI GDPR Developer’s Guide
• Self-Hosted AI and GDPR
• Which AI APIs Are GDPR Compliant?
• AI Privacy Laws by Region
• Best AI Models for Coding Locally (2026)