⚠️ Editorial note: The open source ecosystem in China operates under a distinct institutional framework — characterized by state-led initiatives, intranet-like boundaries, and top-down governance. Readers should be aware that this context differs from the community-driven open source model common in other regions. The term “open source” as used in Chinese media may refer to practices that diverge from the conventional definition.

China Open Source Daily — 2026-11-06

🏛️ Policy & Regulation — The Carve-Out That Runs on an Honour System: Thirty Days Between Announcement No. 48 and the Wuzhen Dialogue

1. The window, stated first because it is the news. The public consultation on MOFCOM and MOST Announcement No. 48 of 2026 — the draft revision of the Catalogue of Technologies Prohibited or Restricted from Export — closes on 2026-11-15. It opened on 2026-10-16, the day the announcement was published, and carried 40 proposed additions across three sectors: 23 AI technologies, 5 biomedical and 12 photovoltaic. Of the 40, 14 are prohibitions (J codes) and 26 are restrictions (X codes). The interim pre-export registration requirement took effect the day of publication, before any comment period had closed, and — in the announcement’s own words — also covers “exports that have already commenced but have not yet been completed.”

2. The clause everything turns on, and the mechanism that sits under it. The announcement states, in the official English rendering carried by the ministries’ own channels, that “technologies that have already entered the public domain through open-source releases, academic publications, or other means will not be subject to these controls.” Read alone, this is the back door this series has been tracking since 2026-10-23: publishing is a lawful exit.

Read against the same document’s registration section, it becomes something more precise and more awkward. The registration requirement lists four triggering transactions, and one of them is not a transaction in the ordinary sense at all:

(3) Changes in the ownership or actual control of domestically held technologies whereby the owner or actual controller becomes an overseas entity.

Fangda Partners’ analysis, published 2026-10-20 and written for clients rather than for readers, draws the conclusion in one sentence — the arrangement “具有类似对’视同出口’行为施加监管的效果,即技术即使未发生物理意义上的跨境,也会因所有权和控制权变化而受到监管”: it functions like a deemed-export rule, in which technology can be regulated even when nothing crosses a border, purely because ownership or control changed. The same note records that the announcement explicitly treats open-sourcing itself as an act of technology export, and warns that “企业在首次开源前必须审慎评估拟开源技术对应的监管要求,而不能直接当作已公开技术处理” — a firm must assess the rule before its first open-source release and cannot treat the technology as already public. The carve-out does not reach back to the act that creates the public domain.

3. The legal gap the carve-out repairs, in the profession’s own words. Fangda’s most useful contribution is not a conclusion but a diagnosis of what the draft is fixing. Under the Regulations on the Administration of Technology Import and Export, there was no provision excluding already-public technology from the export regime at all — with the consequence that “只要是涉及被禁止出口和限制出口的技术,即使已完全公开,其跨境许可、转让理论上仍需申请出口许可证”: even fully public technology would in theory still require an export licence for cross-border licensing and transfer. The dual-use export-control list contains a public-domain exclusion; the technology-import-export regime did not. Announcement No. 48 imports the dual-use logic into AI for the first time, and the firm’s read is that this brings China’s rules into line with how other jurisdictions treat already-public items.

That is a real institutional improvement and it should be recorded as one. A regime that requires a licence to license a paper you have already published is close to unworkable, and closing that gap is what a legal system does when it is being maintained rather than merely extended. The improvement moves the question rather than answering it, however — because the exclusion is triggered by a status the releasing firm asserts about itself, and the register holds that firm’s own account of its own release.

4. Automatic licensing, and the five situations that define the edge of “free.” The draft creates a fast-track path — 自动许可, “automatic licensing,” described by MOFCOM’s spokesperson as intended to operate near-immediately (“基本实现’即报即出’”) but explicitly not an exemption. For restricted AI technologies it applies in five named situations: open sourcing; Chinese firms building and operating their own overseas computing centres; overseas chip tape-out; joint R&D using foreign resources; and participating in international standards-setting.

Two readings of that list are available and both are worth stating. Read generously, it is an enumeration of legitimate purposes the state wishes to protect — a whitelist keeping the ordinary business of an AI lab out of the licensing queue. Read institutionally, it is a map of the activities the state intends to see in advance, and the first item on it is open sourcing. The obligations differ in weight, and Fangda quantifies the asymmetry: an unregistered free-technology contract draws a correction order, a warning, and — if refused — a fine of no more than RMB 50,000 under Article 72 of the Foreign Trade Law; exporting a restricted technology without completing the automatic-licensing process engages the heavier penalties under the technology-import-export regulations. The penalty for failing to declare is not the penalty for exporting; it is the penalty for hiding.

5. The case that has already been through the door. On 2026-10-22, the state-broadcaster-affiliated account Yuyuan Tantian reported that DeepSeek and MiniMax had completed registration under the interim rule and “完成开源和海外授权使用” — completed the open-sourcing and the overseas authorised use. Reuters and Bloomberg both relayed it on 2026-10-23. The same account framed the catalogue’s purpose as “划定边界,而非单纯限制AI技术对外出口” — drawing boundaries rather than simply restricting AI technology exports. Two firms in six days, reported as a smooth procedure rather than a privilege granted, is the strongest available evidence that the instrument is designed to be complied with rather than to bite.

6. Run the third trigger against the documents this series has already read. The registration is live, and its control trigger is a change in actual control. Whether an offshore holding structure, a listing on a non-mainland exchange, or a licence widened in scope until it is economically a transfer counts as such a change is, on the published text, undetermined. Fangda states the same point from the acquirer’s side: “股权比例并非唯一判断因素,少数股权也可能伴随技术控制或使用权的变化” — the equity percentage is not the only factor, and a minority stake can accompany a change in technical control or use rights. This is the same liability question this series examined on 2026-11-04 through four published licence texts from one company, and the same control question it examined on 2026-10-10 through DeepSeek’s limited-partnership structure with concentrated voting rights and a lock-up. None of the three instruments answers the other two, and Chinese AI firms are now operating under all three at once.

7. What the consultation window actually is. Comments go to jinxianmulu@mofcom.gov.cn until 2026-11-15, on the scope, technical thresholds, control measures and facilitation arrangements. The threshold that most needs comment is the one written in capability rather than in compute: the restricted large-model entry covers weight parameters, key checkpoints, underlying source code and training-optimisation technology where training compute reaches 10²⁶ FLOPs, parameters reach 2 trillion, or capabilities reach “the world-leading level” — the last defined as ranking among the global top 10 in public evaluations on aggregate performance, specialised capability, global call volume, safety performance or training efficiency. A threshold that a lab reaches by being ranked is a threshold that moves when someone else ships a model. For a firm whose strategy is to be, in Gao Xi’s phrase, the one that makes AGI arrive sooner, being good is the trigger.

Institutional Analysis & Commentary:

The question this regime answers is not “may we open-source?” It is “who certifies that we did.”

Williamson’s L2/L3 split does the work here, in an unfamiliar direction. At L2 — the formal rules — the draft is doing something careful: it distinguishes prohibited from restricted, imports a public-domain exclusion, writes five automatic-licensing situations, and attaches numeric thresholds. That is the behaviour of a legal system that expects to be complied with, and the Fangda analysis treats it that way. At L3 — governance, who verifies and who decides — the instrument is a filing: the exporter registers, describes the technology, its use and its destination, and proceeds. A filing regime is not a permission regime. It is an information regime, and information regimes run on the accuracy of what the filer writes down.

That is the honest description of what the public-domain carve-out does. It converts is this technology already public? from a question a regulator might investigate into a representation by the releasing firm — one the state has no independent reason to audit and no visible mechanism for auditing. The carve-out does not ask the state to trust the firm’s motives; it asks the state to accept a verifiable fact. A model’s weights are, or are not, downloadable from a named URL on a named date. That fact is cheap to check, and nobody has said who checks it — and in the meantime, by the regulator’s own procedural design, the decision about whether an export control applies is made in the registrant’s queue.

Every regime has to locate the audit somewhere, and this one locates it in the truthfulness of a form. The United States locates it in a licence application adjudicated by the Bureau of Industry and Security; Europe locates a share of it in the open publication that export-control regulations simply do not reach. Announcement No. 48 locates it in a declaration — and the same window that creates that dependency lists open sourcing first among the five purposes granted automatic licensing. The state has decided that Chinese open-weight release is a thing it wants to happen and wants to know about, in that order.

The reward side deserves the same plain statement. A regime that says “publish and you exit” and then makes publication cheap to declare is not a regime trying to stop open-weight release. It is a regime trying to keep a register of it — and a register has a purpose beyond control, which is to know what the country’s AI firms are actually doing abroad. That purpose is legitimate, unglamorous, and plausibly the real one. The industrial-policy reading is right there in the ministry’s own framing: boundaries, not bans.

Where it gets sharp is the second-order condition. A regime that prices open publication at zero and closed licensing at a licence fee is a subsidy to openness, and subsidies get used. DeepSeek’s V4.1-Flash under plain MIT; Qwen’s 460-plus models and 4 billion cumulative downloads; MiniMax publishing not only weights but the failed experiments, the data and the tooling modifications, in Feng Wen’s account at Wuzhen — under this instrument each of those acts is simultaneously a strategic move, an engineering choice, and a legal event. Some require a filing. All now sit inside a category the state registers.

One perspective, not a verdict. The checkable question is narrow and answerable: does the first published set of registration statistics break filings down by type — open-source release, overseas licence, control change, self-built compute abroad — or only by sector? A breakdown by type would tell an outside reader what China’s AI firms are actually exporting, which is precisely the figure everyone currently estimates from download counts and funding rounds. A single aggregate would not, and the aggregate is the cheaper thing to publish. Which one appears is the institutional fact to watch, and the mailbox is open for another nine days.


🏛️ Policy & Regulation — Model Tiering Arrives as a Remedy Before It Arrives as a Rule: What Three Labs Said at Wuzhen About the Regime That Is Coming

1. The setting, and the moderator’s framing. The “AI Wuzhen Dialogue” (人工智能乌镇对话) ran on the afternoon of 2026-11-03 inside the main forum of the 2026 World Internet Conference Wuzhen Summit — the first edition whose theme places 开源 (open source) at its centre, with roughly 1,800 delegates from 138 countries and regions in attendance. It was chaired by Wang Jian (王坚), academician of the Chinese Academy of Engineering, director of Zhejiang Lab, and chief director of the WIC’s AI committee. The three firms on stage were DeepSeek, Alibaba Qwen and MiniMax, joined by John Higgins, chair of the International Association for AI Governance.

Two of Wang Jian’s opening sentences structure the entire hour:

“在软件时代,中国更多的是开源思想的受益者;今天在人工智能时代的开源路径上,中国变成了非常重要的贡献者、引领者” — in the software era China was mostly a beneficiary of the idea of open source; on the AI-era open-source path China has become an important contributor and leader.

“开源追求开放,但开放的前提是可信、可控、可治理” — open source pursues openness, but the premise of openness is trustworthiness, controllability and governability.

2. Gao Xi (DeepSeek) on why, and on what it costs. “很多人问,DeepSeek为什么要把辛辛苦苦研究而出的成果开源出来?其实答案很简单,就是为了实现AGI” — the answer is simple: to make AGI arrive. The competitive framing is deliberately not about rivals: “对大模型企业来说,在人工智能竞赛中,对手不是其他公司,而是时间” — the opponent is not another company, it is time. And the cost is stated in the present tense: “做开源,确实有可能会放大人工智能导致的安全风险,所以对DeepSeek来说,有责任去避免安全问题被扩大” — open source may indeed amplify AI-related safety risks, and DeepSeek has a responsibility to prevent that amplification. On remedies, Gao said his colleagues were researching approaches to these problems “including approaches such as model tiering” (比如模型分级).

3. Liu Dayiheng (Alibaba Qwen) on the two structural defects. Qwen’s open-weight record, in his own figures: 460-plus models, over 4 billion cumulative downloads, over 330,000 derivative models since the first 7B model three years ago. His diagnosis is the most precise institutional statement made on that stage, and it separates two things the public debate routinely conflates:

First, non-recallability (不可撤回性). Once released and widely propagated, a model with a discovered defect cannot be repaired in place the way a cloud-hosted model can be patched centrally on the provider’s servers.

Second, the absence of an external perimeter (缺少外生的护栏). Open-model safety depends on internal post-training capability, and a third party holding the weights can look for ways around the guardrails.

His conclusion is the one that matters: “仅仅靠对模型做限制,很难从根本上解决治理问题,还需更多的法律法规、行业规范和配套基础设施共同发力” — restricting models alone cannot fundamentally solve governance; laws, industry norms and supporting infrastructure must work together. Then the sentence that reframes the conversation: “开源和安全不是简单的对立关系,治理才是那个需要被认真对待的变量” — open source and safety are not a simple opposition; governance is the variable that must be taken seriously. He gave the counter-argument its due too: open models are white-box, more researchers can attack them, and defensive capability may grow with offensive capability — “两三年前我们会认为一个1.5B这种很小规模尺寸的模型做不了什么事情,但在新的后训练技术面前,它可以去解锁这个模型的能力。”

4. Feng Wen (MiniMax) on what “open source” has expanded to contain. MiniMax’s practice publishes more than weights: “除了开源不同模态的模型,还会从工程实践角度分享研发过程中的经验,包括遇到的问题、开展的实验、相关数据,以及对工具所做的改造” — the problems encountered, the experiments run, the associated data, and the modifications made to tooling. One concrete return on that: after open-sourcing a video-generation model, chip vendors, independent developers and content creators adapted it, some teams re-trained and optimised inference, and the result was a five-second video generated in under two seconds — beyond what the originating team had anticipated. Her industrial claim: security-conscious firms choose open models because they can fine-tune and deploy privately, which makes an open weight a more自主可控 (self-controllable) choice rather than a less safe one.

5. Higgins on the missing institution. The international-governance participant’s contribution was an absence stated as a requirement: providers in the open-source field must overcome the safety problem, and the open-source world needs community co-construction and shared responsibility, with joint evaluation and testing and a cross-border certification system to give the global market a trusted basis. In the SCMP rendering: the real problem is not technology but trust, and the challenge for firms is how to demonstrate confidence.

6. The regulator’s parallel statement, made the same day in a different room. At a Beijing briefing on 2026-11-03, Wang Ruomeng, a senior official at the National Development and Reform Commission, said — per Bloomberg — that China “will accelerate legislation to support core AI enablers — data, computing power and algorithms — and regulate key application scenarios, while establishing safety standards and ethical guidelines covering the full life cycle of AI.” Developers and operators should strengthen risk management and disclose safety risks, incidents and vulnerabilities. The same reporting describes an unspecified transition period, official guidance, and sector-specific regulatory sandboxes, and states that open AI development remains encouraged through the 15th Five-Year Plan to 2030. No final text, effective date or penalty schedule was published. It is a direction, and it should be recorded as one.

Institutional Analysis & Commentary:

The most consequential sentence on that stage was not about safety. It was about who pays for it.

Put the two diagnoses side by side. Gao Xi said open source may amplify risk and that DeepSeek is responsible for preventing that amplification, with model tiering as live research. Liu Dayiheng said open models are non-recallable and lack an external perimeter, and that restriction alone will not solve governance — laws, industry norms and infrastructure must do the work together. Higgins said the world needs shared evaluation, testing and cross-border certification.

Every item on that list is an obligation of the releasing firm, plus a facility that does not exist. Non-recallability is a property of the artefact, not a fault of the releaser: nobody in any jurisdiction can un-publish a weight file. The absence of an external perimeter is equally structural. What can reduce the resulting harm is a test that somebody runs, a report that somebody reads, and a certification someone accepts — and in the open-weight case none of those can be performed by the firm whose model is being tested without becoming an assertion rather than a verification. That is the whole institutional gap, and it is not a Chinese gap. It is the gap between a private cost (train the model, publish it, document it) and a public good (a security commons of evaluations, incident reporting and independent audit), and no ministry in any country has yet appropriated a budget line for it.

That is what makes the NDRC statement more interesting than it first appears. Disclosure of safety risks, incidents and vulnerabilities is a duty that can be written into law, and once written it produces a register of incidents — the one asset an evaluation commons needs and cannot manufacture, because it is the only way an outside party learns what an open model did after it left the building. But a register of disclosures is not an evaluation and not a certification; it is a filing, and it runs on the same honour system as Announcement No. 48’s public-domain carve-out. Two instruments, three weeks apart, both locate the audit in what the firm asserts. That is not a scandal. It is the cheapest available design, and its cost lands in a specific place: on the firms that can afford to run their own evaluations and would rather be certified by someone else.

There is a coordination problem underneath, and Liu Dayiheng named it without naming it. If one lab funds an independent evaluation body and publishes incident data while its competitors do not, it absorbs a cost that improves everyone’s safety — including rivals’ — while rivals ship faster. Every lab privately wants the certification Higgins described; no lab privately wants to be first to pay for it. That is the standard structure of an under-supplied public good, and the standard remedies are a mandatory standard, a state-funded facility, or a coalition with an exclusion rule for free riders. The NDRC’s mention of transition period, guidance and regulatory sandboxes is the state choosing, so far, the softest of the three.

Wang Jian’s framing should be marked as a claim rather than a finding. In the software era China benefited; in the AI era China contributes and leads. The contribution half is documented — Qwen’s 460-plus models, DeepSeek’s MIT-licensed releases, MiniMax’s engineering disclosures, and the OpenRouter weekly call rankings cited throughout Chinese coverage. The leadership half rests on download volume and call volume — quantities that measure adoption, not governance. A country can lead in release volume and lag in the institutions that make release verifiable; those are different scoreboards, and this series has watched them be conflated for weeks.

One detail belongs in the record because it is where “recall” and “withdrawal” met. Liu Dayiheng’s 不可撤回 — non-revocable, non-withdrawable — is not the property examined in this series on 2026-11-04, where the question was whether a licence grant can be revoked. The first is a technical impossibility: you cannot collect a downloaded weight file. The second is a legal restraint the grantor chose not to impose. The first makes the second expensive to promise and impossible to enforce, and neither the custom model licences nor the open-source licences surveyed this week address what happens when a defect is found in a model four million people have already cloned. That is the question no document published in the last three weeks has answered.

Checkable questions, all cheap.

(i) When the NDRC lifecycle framework is published, does incident disclosure attach to open-weight releases specifically, and does the reporting duty fall on the releasing developer, on downstream deployers, or on both? The answer determines whether the register describes the artefact or the deployment.

(ii) Does the transition period and the sandbox regime name an independent evaluation body — a third-party laboratory, a standards institute, an international certification scheme — or only an internal compliance function? Higgins asked for the first. A framework answering with the second has outsourced the commons back to the firms.

(iii) Does DeepSeek publish the model-tiering work Gao Xi described as in progress? Model tiering — graded capability release, presumably with the most dangerous capabilities withheld from open weights — is the most plausible technical answer to non-recallability available today. Published as a method, it is a contribution to everyone’s safety. Kept as an internal practice, it is a competitive asset every lab must re-derive. Which one it becomes will say more about this ecosystem’s governance trajectory than any summit theme.


🔍 Commentary — A Security Commons That Is Nobody’s Facility

First: the exporter cannot be the auditor, and the draft has made that structural. Announcement No. 48 asks a firm to declare whether its technology is already public; the NDRC framework asks the same firm to declare incidents. Both are reasonable, both are cheap, and both are self-assessments by the party with the strongest interest in the outcome. The alternative — an inspectorate — costs money, takes years, and would be slower to build than the models it would inspect. So the honour system is not a mistake; it is the only design available at current speed. But an honour system requires an honour code, and an honour code requires a profession with a reputation to lose. Chinese open-weight labs are building that profession in public, which is more than most industries manage.

Second, the reciprocal question, asked plainly because this series exists to ask it of everyone. The United States, through the NSA-CISA-FBI advisory this series documented on 2026-09-28, treats Chinese open weights as a proliferation problem to be named. The European Union’s AI Act places its heaviest obligations on general-purpose models above a compute threshold, with an open-source exemption narrower than is often assumed. China’s draft places a publication carve-out deeper than either — it does not merely exempt open models, it declares publication to be the lawful route out of the regime. Each jurisdiction is making a different social choice about whether openness is a risk to be managed or a public good to be produced. None of the three has funded the evaluation infrastructure that would let the choice be tested, and the jurisdiction that builds it first will set the standard the others argue with.

Third: what the week’s most-read Chinese coverage got right, and what it left out. The Wuzhen dialogue’s coverage converged on a single line — Liu Dayiheng’s “治理才是那个需要被认真对待的变量” — which is genuinely the correct sentence. The omission is that governance is not a variable; it is a set of costs and a set of beneficiaries, and the coverage did not say who bears them. A mandatory evaluation regime raises the fixed cost of release. That cost is trivial for Alibaba and DeepSeek and non-trivial for a two-person team in Chengdu that has fine-tuned its way to a useful model. Every safety regime a large lab can afford is a barrier to entry for a small one, and the labs writing the code of conduct are not the labs the code of conduct burdens. This is the oldest problem in the political economy of standard-setting, and it does not become less true when the standard is written for honest reasons.

Fourth, and this is the one worth keeping for the file. The word at the centre of the week is 分级 — grading, tiering. Gao Xi offered it as a technical remedy. The MOFCOM catalogue uses it as a mechanism (J codes against X codes, prohibitions against restrictions, numeric thresholds instead of categories). The NDRC framework implies it (transition period, sandboxes, sector-specific rules). A national standard this series documented on 2026-09-19 already grades large models on an assessment scale. Four institutions reached for the same instrument within four weeks. Tiering is what a state reaches for when a technology’s risk is continuous and its politics is binary: it converts a yes-or-no question into a classification question. That is progress over a ban, and it is a real achievement. It is also the point at which the classification decision — who is in which tier, and who decides — becomes the entire game. Watch the classification body, not the tier definitions. The definitions will be published; the body’s discretion is where the policy actually lives.


Sources: MOFCOM & MOST — Announcement No. 48 of 2026, draft revision of the Catalogue of Technologies Prohibited or Restricted from Export (published 2026-10-16; consultation to 2026-11-15) · MOFCOM spokesperson Q&A on the catalogue revision (2026-10-16): 14 prohibitions, 26 restrictions, precisely limited AI scope, automatic licensing · Xinhua — two ministries open public consultation on the catalogue revision (2026-10-16) · Fangda Partners — reading the catalogue revision and the interim registration requirement (2026-10-20): deemed-export effect of control changes, public-domain exclusion repairing the Regulations’ silence, automatic licensing distinguished from exemption, Article 72’s RMB 50,000 ceiling · Geopolitechs — Announcement No. 48 with Annex 1 AI control points (10²⁶ FLOPs / 2 trillion parameters / global top 10; agents at 20 subtasks and 200 tools; 800 GB/s interconnect) · FourWeekMBA — DeepSeek and MiniMax complete registration and overseas licensing, per Yuyuan Tantian 2026-10-22, relayed by Jiemian and Bloomberg · ChinaTalk — MOFCOM’s New Rules and the Future of Chinese AI (2026-10-21): 14 J codes, 26 X codes, licence adjudication between central and provincial authorities · Xinhua — reading the AI open-source-ecosystem keywords out of the Wuzhen dialogue (2026-11-04): Wang Jian on beneficiaries and contributors; openness premised on trustworthiness, controllability, governability · Sina Finance / The Paper — the three top model firms on one stage at Wuzhen (2026-11-03): Liu Dayiheng on non-recallability and the missing external perimeter; Gao Xi on time as the competitor; Feng Wen on publishing experiments and data · China News Service / Guangming — AI Wuzhen Dialogue: the question of open source’s value, the way of governance, the wish for inclusiveness (2026-11-04) · Geopolitechs — DeepSeek researcher explains why open source matters at Wuzhen (full English rendering of the roundtable) · SCMP — DeepSeek engineer defends the open-source push despite higher risks (2026-11-04): 1,800 delegates from 138 countries; domestic registration for DeepSeek and MiniMax · Guangming — from “buckle up” to “human-centred”: the Wuzhen summit on AI safety (2026-11-04): Joe Tsai on open source as the only route to普及; ZTE on full-lifecycle governance · ExplainX — what the NDRC actually said (Wang Ruomeng, Beijing briefing 2026-11-03, per Bloomberg): lifecycle safety standards, incident disclosure as “should”, transition period and sector sandboxes, no final text or penalty schedule · The Paper — DeepSeek consolidates Harness, Web, App and Platform teams into a Product Division under Cui Tianyi (2026-10-16)