Beijing’s Ministry of Commerce and Ministry of Science and Technology unveiled a draft revision to the country’s export‑control catalogue on 16 October 2026, targeting 40 strategic technologies—including AI foundation‑model weights, high‑performance chip design know‑how and certain biotech processes. The proposal, announced as Announcement No. 48, immediately imposes a temporary pre‑export registration requirement while a 30‑day public comment period runs until 15 November 2026. If adopted, the rules would be the first Chinese outbound controls that explicitly cover artificial‑intelligence software, model checkpoints and related intellectual property.
From Consultation to Draft Law
During the summer of 2026, MOFCOM held informal talks with leading domestic AI and semiconductor firms—Alibaba, ByteDance, Zhipu AI and others—to map the perimeter of software and IP that could be treated as export‑controlled. The outcome was a draft annex that adds 40 categories to the existing “Catalogue of Technologies Prohibited or Restricted from Export.” Fourteen of those categories are outright prohibited; the remaining twenty‑six require a state licence.
The draft takes effect immediately for pre‑export registration: any Chinese exporter must file a notice with MOFCOM for pending or future transfers of the listed items, even if the transaction was already underway before 16 October. The registration rule is interim; a full set of implementation regulations will follow the public‑comment deadline.
What Technologies Are Covered?
Artificial‑Intelligence Stack (23 categories)
The AI portion of the draft is the most expansive. It reaches into three layers of the ecosystem:
| Category | Example items |
|---|---|
| Foundation‑model weights and checkpoints | Full‑size model parameters, training‑state files |
| Model architecture and optimization techniques | Neural‑network designs, sparsity or quantisation methods that boost compute efficiency |
| AI‑chip design know‑how (≤ 14 nm, 2.5 D/3 D packaging) | Layouts, mask data, advanced memory‑cell designs |
| AI‑assisted vulnerability‑discovery tools | Software that automatically scans code for exploits |
| High‑performance deep‑fake detection tech (≥ 95 % accuracy) | Algorithms used for cross‑border network‑security filtering |
| Autonomous malicious‑code generators | Systems that can create weaponised code without human input (prohibited) |
Quantitative Triggers for Large‑Language‑Model Controls
Item 266504X applies the export restriction when any of the following thresholds is met:
- Compute volume – training compute of ≥ 10^26 FLOPs.
- Parameter count – models with ≥ 2 trillion parameters.
- Global ranking – a model placed in the top‑10 worldwide on recognized public benchmarks covering overall performance, safety or domain‑specific tasks.
- Data sensitivity – models trained on China‑designated “important data” (under the Data Security Law) that are offered as services abroad. (Models with < 0.1 % uncleaned sensitive data that never appear in outputs are exempt.)
If a model satisfies any one of these criteria, it falls under the “restricted” regime and must obtain a licence before its weights, source code or deployment rights can be transferred overseas.
Photovoltaics and Solar Energy (12 categories)
The draft also clamps down on advanced solar‑manufacturing techniques, including polysilicon processing, wafer‑cutting and perovskite tandem‑module technology. Perovskite module tech is listed as prohibited, meaning any export of the underlying process knowledge is barred outright.
Biomedicine and Biotechnology (5 categories)
The biotech slice covers cell‑ and gene‑therapy platforms, small‑nucleic‑acid drugs and certain pathogen‑related technologies. These are treated similarly to the AI items: some are prohibited, others require licensing.
Prohibited vs. Restricted: The Practical Difference
- Prohibited items cannot be exported under any circumstance. This includes autonomous AI systems that generate malicious code and ultra‑accurate deep‑fake detection tools used for cross‑border network security.
- Restricted items may be exported if the exporter obtains a licence from MOFCOM. The draft mentions an “automatic licensing” pathway for companies that pre‑register and meet certain conditions (e.g., deploying a model on a corporate‑owned cloud abroad or releasing an open‑weight version).
Carve‑Outs and Service‑Only Exemptions
The regulation distinguishes between the transfer of intellectual property and the provision of services built on that IP:
- Public‑domain and open‑source exemption – Anything already released in the public domain through academic papers or open‑source repositories before the announcement is fully exempt.
- Service access is not an export – Offering a model via an API, SaaS platform, web portal or mobile app does not count as exporting the technology. Only the transfer of weight files, source code, or licensing rights for local deployment abroad triggers the control.
- Edge‑device deployment exemption – Lightweight, on‑device inference that does not involve transferring the full model weights is excluded.
These carve‑outs aim to preserve China’s participation in global AI research while tightening control over high‑value, high‑risk assets.
Who Will Feel the Impact?
The draft’s language is company‑agnostic; it does not name any specific model or developer. Nonetheless, analysts have flagged several home‑grown models that appear to cross the quantitative thresholds:
- Alibaba’s Qwen 3.8‑Max – rumored to exceed 2 trillion parameters.
- Moonshot AI’s Kimi K3 – reportedly ranks among the top‑10 globally on several benchmark suites.
If the final rules adopt the draft’s thresholds, these models would require a licence for any overseas transfer of their weights or architecture. The same applies to any Chinese firm developing AI chips at 14 nm or below, a capability that several domestic semiconductor companies have announced in recent months.
Geopolitical Context: A “Bidirectional Squeeze”
The United States has long used inbound export controls to restrict the flow of high‑end GPUs and other hardware into China. Beijing’s move represents a reversal: outbound controls that limit Chinese AI software and IP from reaching foreign users or partners. Observers describe the development as a “bidirectional squeeze” in AI governance, where both sides of the trade flow are increasingly regulated.
A MOFCOM spokesperson framed the policy as a safeguard: “We must prevent misuse and abuse, ensure AI remains under human control, and protect national security, while continuing to encourage open‑source collaboration and international technological sharing.” The language underscores a dual objective—security tightening without choking the country’s participation in global AI ecosystems.
Ambiguities and Open Questions
Global‑Top‑10 Benchmark Definition
The draft requires a model to be in the “top‑10 globally” on “mainstream public evaluations,” but it does not specify which benchmarks count (e.g., HELM, BIG‑Bench, or proprietary leaderboards). The lack of clarity could lead to divergent interpretations by firms and regulators.
Future Open‑Weight Releases
Existing open‑source releases are grandfathered, yet the draft introduces an “automatic licensing” mechanism for future open‑weight models. The exact criteria and processing timelines for this pathway remain undocumented, leaving companies uncertain about how quickly they can comply.
Implementation Burden of Interim Registrations
Exporters must now file a registration for any pending cross‑border transaction involving the listed items. MOFCOM has not disclosed processing times or approval rates for these interim filings, raising concerns about potential bottlenecks for ongoing deals.
Unnamed Model Impact
Media outlets have speculated that specific Chinese models exceed the 2‑trillion‑parameter threshold, but the draft itself refrains from naming any. This creates a gray area: firms must assess whether their own models meet the criteria without explicit guidance, potentially leading to over‑cautious self‑restriction or inadvertent violations.
What Comes Next?
The public consultation runs until 15 November 2026. Stakeholders—including AI labs, semiconductor firms and foreign partners—can submit comments on the draft’s scope, thresholds and licensing procedures. After the comment period, MOFCOM and MOST will review feedback, finalize the catalogue revisions and publish detailed implementation rules. Until then, the temporary registration requirement remains in force, and any cross‑border transfer of the covered technologies should be treated as a regulated activity.
The forthcoming final rules will shape how Chinese AI firms navigate the global market. If the licensing process proves swift and transparent, Chinese companies may continue to export cutting‑edge models under licences, preserving their competitive edge. Conversely, a restrictive or opaque regime could push firms toward more aggressive open‑source strategies or encourage relocation of development activities to jurisdictions with fewer export constraints.